Breeding performance prediction system based on intelligent sheep breeding platform

Abstract

The invention discloses a breeding performance prediction system based on an intelligent sheep breeding platform, and belongs to the technical field of animal husbandry informatization. The intelligent breeding platform is designed, standardized collection, storage and intelligent management of sheep full-life-cycle data are achieved, and a solid data foundation is laid for follow-up research; secondly, innovatively combining ensemble learning (AdaBoost, GBRT and the like), machine learning (SVR, KNN and the like) and a sorting learning algorithm, and constructing a multi-level breeding sheep breeding performance prediction model system; particularly, a subjective and objective combination weighting method based on AHP-PCA is provided, and the prediction precision of the model is remarkably improved through feature reconstruction; in addition, an Achimedes optimization algorithm (AOA) is introduced into the field of breeding prediction for the first time, and adaptive search of hyper-parameters is realized by using a physical simulation mechanism of the AOA, so that the accuracy, efficiency and robustness of the model are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of information technology in animal husbandry, specifically to a breeding performance prediction system based on a smart sheep farming platform. Background Technology

[0002] With the advancement of livestock modernization, accurate prediction of sheep breeding capacity has become a key technology for improving breeding efficiency and optimizing population structure. Traditional breeding methods rely on manual experience or basic statistical models, which suffer from low prediction accuracy and insufficient efficiency. Although biogenetics has made progress in breeding technology, its widespread application in regions with lagging information technology still faces challenges such as high costs and technical barriers.

[0003] In traditional sheep farming, growth data is primarily recorded manually, leading to problems such as data disorganization, storage difficulties, and lack of traceability. The lack of standardized data management platforms in remote areas prevents researchers from efficiently utilizing data, hindering in-depth breeding research.

[0004] In recent years, machine learning techniques have been increasingly applied in agriculture, particularly demonstrating advantages in animal genome selection and phenotypic data analysis. For example, ensemble learning algorithms (such as random forests and AdaBoost) outperform traditional statistical methods in predicting the Comprehensive Breeding Value (GEBV); regression models (such as support vector regression and K-nearest neighbors) have been used to analyze the association between sheep weight, body size, and reproductive traits. Ranking learning algorithms show promise in multi-objective ranking and complex prediction tasks, but their application in animal husbandry has not yet been fully explored.

[0005] Although traditional breeding methods and some machine learning techniques have been applied in animal husbandry, significant limitations remain. First, traditional methods relying on human experience or basic statistical models struggle to fully uncover potential patterns in multi-source data, resulting in insufficient prediction accuracy and failing to meet the demands of precision breeding. Second, existing algorithms have limited generalization ability when dealing with complex breeding traits, and single models often struggle to consistently capture the nonlinear relationship between sheep growth performance and breeding performance. Furthermore, the low level of intelligence in data collection and analysis, coupled with the lack of a unified standardized platform, leads to inefficient data integration and utilization, particularly in remote pastoral areas. These problems severely restrict the improvement of breeding efficiency and the sustainable development of modern animal husbandry. Therefore, there is an urgent need to develop more efficient data management tools and intelligent prediction models to achieve a more scientific and precise breeding process. Summary of the Invention

[0006] The technical problem to be solved by this invention is: how to solve the problems of insufficient prediction accuracy and low efficiency of data integration and utilization in traditional methods that rely on human experience or basic statistical models, and to provide a breeding performance prediction system based on a smart sheep breeding platform.

[0007] This invention solves the aforementioned technical problems through the following technical solution: The invention includes a smart breeding platform, a data acquisition module, a data processing and analysis module, a machine learning prediction module, a ranking and learning optimization module, and a user interface module. The smart breeding platform is developed based on the PHP framework, adopts a B / S architecture, is deployed on the Alibaba Cloud platform, and supports multi-level user access control. The data acquisition module collects sheep-related data and stores the standardized data in a MySQL database. The data processing and analysis module performs data preprocessing and analysis. The machine learning prediction module predicts the comprehensive breeding value of sheep using a machine learning model, obtaining the predicted breeding performance results. The ranking and learning optimization module retrains the machine learning model to improve accuracy using AHP-PCA combined weighting or AOA optimization. The user interface module displays sheep growth trends and breeding performance prediction results through charts.

[0008] Furthermore, the intelligent livestock farming platform includes a system configuration module, a basic data management module, a feed formulation management module, a pasture management module, a treatment and disease prevention module, a traceability system module, a breeding management module, a storage management module, a data statistical analysis module, and a data recording module. The system configuration module maintains and manages the basic settings and user permissions of the intelligent livestock farming platform. The basic data management module maintains and manages the core basic information within the intelligent livestock farming platform. The feed formulation management module maintains and manages the entire feed formulation process. The pasture management module comprehensively manages livestock information and related operation records within the pasture. The treatment and disease prevention module includes modules for treatment records, disease prevention records, drug records, disease records, and prescription records. The traceability system module comprehensively records and manages the entire process of sheep from acquisition to delivery. The breeding management module comprehensively records and manages the livestock breeding process. The storage management module manages warehouse and inventory information within the pasture. The data statistical analysis module performs statistical analysis on various data within the pasture. The data recording module comprehensively records various data within the pasture.

[0009] Furthermore, in the data acquisition module, for rams, the acquired data includes growth performance indicators, such as body weight, height, length, chest circumference, and cannon bone circumference; for ewes, in addition to the growth performance indicators, reproductive performance indicators are also acquired, such as number of lambs, reproductive survival rate, and twinning rate.

[0010] Furthermore, the data processing and analysis module includes a data preprocessing unit and a statistical analysis unit. The data preprocessing unit standardizes and complies with data formats, removes outliers, corrects errors, and eliminates duplicate records. Through entity recognition, redundancy elimination, and conflict resolution, it aggregates data from different sources into a single dataset and compresses the dataset. The statistical analysis unit performs statistical analysis on the individual sheep phenotypic characteristic data cleaned by the data preprocessing unit. Specifically, Pearson correlation analysis is used to calculate the correlation coefficients between growth performance indicators, reproductive performance indicators, and the overall breeding value to identify the main characteristics affecting sheep breeding performance. Then, path coefficient analysis is used to quantitatively decompose the direct and indirect effects of each characteristic on the overall breeding value, obtaining the causal relationship structure between characteristics. This structure guides the optimization of feature weighting and model input variables, thereby improving the accuracy and stability of the subsequent breeding performance prediction model.

[0011] Furthermore, in the machine learning prediction module, the machine learning model includes a ram model and a ewe model;

[0012] Ram models include AdaBoost, GBRT, RF, and Extra-Trees;

[0013] During runtime, the optimal model is selected from AdaBoost, GBRT, RF, and Extra-Trees based on performance requirements and data characteristics. In the model building phase, cross-validation and performance evaluation are performed on each candidate ram model. The optimal model is determined by comparing their coefficient of determination R², root mean square error RMSE, and mean absolute error MAE.

[0014] During the formal prediction phase, the optimal model is invoked to predict the comprehensive breeding value. The model input is the growth performance index of the breeding ram, and the output data is the comprehensive breeding value of the corresponding individual breeding ram, which is used to represent the genetic breeding potential of the corresponding individual.

[0015] Breeding ewe models include SVR, KNN, LR, and DTR;

[0016] During runtime, the optimal model is selected from SVR, KNN, LR, and DTR based on performance requirements and data characteristics. In the model building phase, cross-validation and performance evaluation are performed on each candidate ewe model. The optimal model is determined by comparing its coefficient of determination R², root mean square error RMSE, and mean absolute error MAE.

[0017] During the formal prediction phase, the optimal model is invoked to predict the comprehensive breeding value. The model input consists of the growth performance index and reproductive performance index of the breeding ewe, and the output data is the comprehensive breeding value of the corresponding individual breeding ewe, which is used to represent the genetic breeding potential of the corresponding individual.

[0018] Furthermore, in the ranking learning optimization module, the specific process of optimization using AHP-PCA combined weighting is as follows:

[0019] S11: Data Preprocessing

[0020] The collected growth performance indicators and reproductive performance indicators were standardized to obtain a standardized feature matrix;

[0021] S12: Subjective Empowerment

[0022] Based on the experience and knowledge of sheep breeding experts, pairwise comparison matrices A for characteristic indicators were constructed for breeding rams and breeding ewes respectively. For breeding rams, the indicators included weight, height, length, chest circumference, and cannon bone circumference; for breeding ewes, the indicators included weight, height, length, chest circumference, cannon bone circumference, number of lambs born, reproductive survival rate, and twinning rate. The relative importance of each indicator was scored using the Saaty scaling method, the largest eigenvalue and corresponding eigenvector of the matrix were solved, and the matrix was normalized to obtain the AHP weight vector. The consistency ratio CR was calculated. When CR < 0.10, the matrix was considered to be consistent; otherwise, the evaluation was returned to the next step.

[0023] S13: Objective Empowerment

[0024] Principal component analysis is performed on the standardized feature matrix to calculate the eigenvalues ​​and loading matrix. The top m principal components with a cumulative variance contribution rate of not less than 85% are selected, and the comprehensive contribution value of each original feature is calculated and normalized to obtain the PCA weight vector.

[0025] S14: Integration of Subjective and Objective Weights

[0026] Set the subjective and objective fusion coefficient α∈[0,1], and use the linear weighting method to fuse the AHP and PCA weight vectors based on the subjective and objective fusion coefficients to calculate the combined weight vector;

[0027] S15: Feature Reconstruction

[0028] Based on the combined weight vector, the standardized feature matrix is ​​weighted to calculate the weighted reconstructed feature matrix;

[0029] S16: Model Training and Validation

[0030] The weighted reconstructed feature matrix and the corresponding comprehensive breeding value were used as input and output samples to train various ram and ewe models. K-fold cross-validation was used to evaluate the performance of each model and select the optimal model.

[0031] S17: Performance Evaluation and Threshold Determination

[0032] Compare the changes in model performance before and after the AHP-PCA combined weighting optimization. When the R² increase value ΔR² ≥ 0.05 or the RMSE reduction rate ≥ 3%, the optimization is deemed effective and the current weighting scheme is fixed. If the threshold is not reached, return to step S12 or S13 to readjust the weights.

[0033] S18: Online Deployment and Periodic Retraining

[0034] During operation, steps S11 to S17 are executed periodically to update the model. When the number of new samples exceeds the set threshold or the time interval exceeds three months, the retraining process is automatically started.

[0035] S19: Exception and Robust Handling

[0036] When the consistency of the pairwise comparison matrix A fails, revert to the PCA single-weighting mode; when the cumulative variance contribution rate of PCA is less than 60%, suggest expanding the sample or introducing new features to improve the information content; remove invalid features with zero or near-zero variance.

[0037] S110: Result Output and Storage

[0038] The output results, including the combined weight vector, the optimal model type and parameters, and the model performance indicators, are stored in the database for subsequent breeding performance prediction and model iteration.

[0039] Furthermore, in the ranking learning optimization module, the specific process of the AOA optimization method is as follows:

[0040] S21: Definition and Coding of the Optimization Problem

[0041] Determine the set of variables P to be adjusted in this AOA optimization, which includes the set of hyperparameters of the machine learning model; map P to a continuous real number vector representation, define the upper and lower bounds for each dimension, and finally define the fitness function;

[0042] S22: Initialize population and individual physical attributes

[0043] Randomly generate N candidate solution position vectors, and uniformly sample each position vector within the corresponding range. The candidate solution is the individual. Initialize physical properties for each individual to drive subsequent position updates. The physical properties include density, volume and acceleration. Calculate the fitness of each initial individual and record the current best individual and its fitness.

[0044] S23: Start the main loop and begin iteration.

[0045] Enter the iterative loop until the stopping condition is met;

[0046] S24: Update physical properties

[0047] For each individual k, update the density and volume based on its current fitness and the population's best fitness: if the individual has a better fitness than the population average, adjust its density or volume in a favorable direction; then calculate the individual acceleration.

[0048] S25: Determine the search phase of the current iteration.

[0049] Set the stage probability or transition function according to the iteration progress, and use the transition operator or probability threshold in AOA to determine whether this is "collision / random movement" or "acceleration towards the optimal state";

[0050] S26: Location Update

[0051] If in the exploration phase, update the individual position according to the exploration strategy; if in the utilization phase, update the position towards the optimal individual according to the acceleration. Perform boundary checks on the updated position and map it back to the legal domain.

[0052] S27: Fitness Assessment and Population Replacement

[0053] For each new position, the actual parameters are decoded. In the training / validation process, the corresponding model is built and the fitness is calculated. If the fitness is better than the original fitness of the individual, the original individual is replaced and the global optimal individual is updated. The global optimal individual is the global optimal solution.

[0054] S28: Local Search Enhancement and Fine-tuning

[0055] When there is no improvement for several consecutive generations or the preset early convergence criterion is reached, a local search is triggered. If a better solution is obtained by using the local search, the global optimal solution is replaced.

[0056] S29: Stopping Condition Judgment

[0057] When the fitness of the globally optimal individual improves less than a threshold over consecutive generations. If the iteration is terminated, otherwise return to step S24 to continue the iteration. When the iteration is terminated, the final optimal parameters, the corresponding model performance index, and the convergence history curve are output.

[0058] S210: Result Validation, Deployment, and Rollback Strategies

[0059] Final validation is performed on an independent test set, and R², RMSE, and MAE metrics are recorded. If the preset performance target is met, the model will be written into the model library and deployed as online prediction parameters. If the preset performance target is not met, alternative strategies are initiated: using the AHP-PCA combined weighting results as a baseline, or expanding the search space, increasing the population size and number of iterations, and then repeating steps S21 to S29.

[0060] Compared with existing technologies, this invention has the following advantages: Firstly, it designs an intelligent breeding platform, realizing standardized collection, storage, and intelligent management of sheep life-cycle data, laying a solid data foundation for subsequent research. Secondly, it innovatively combines ensemble learning (AdaBoost, GBRT, etc.), machine learning (SVR, KNN, etc.), and ranking learning algorithms to construct a multi-level sheep breeding performance prediction model system. In particular, it proposes a subjective-objective combination weighting method based on AHP-PCA, significantly improving the model's prediction accuracy through feature reconstruction. Furthermore, it introduces the Archimedes optimization algorithm (AOA) into the field of breeding prediction for the first time, utilizing its physical simulation mechanism to achieve adaptive hyperparameter search, resulting in significant improvements in accuracy, efficiency, and robustness. Attached Figure Description

[0061] Figure 1 This is a schematic diagram illustrating the basic functions and requirements of the smart aquaculture platform in this embodiment of the invention;

[0062] Figure 2 This is a schematic diagram of the functional interface division of the intelligent aquaculture platform in an embodiment of the present invention;

[0063] Figure 3 This is a simplified schematic diagram of the structure of a three-dimensional sheep body ruler in an embodiment of the present invention;

[0064] Figure 4 This is an example point cloud diagram of sheep body size measurement in an embodiment of the present invention;

[0065] Figure 5 This is an example of a body size data recording interface in the ranch management interface of this invention.

[0066] Figure 6 This is an example of the main console interface of the intelligent sheep farming platform in this embodiment of the invention;

[0067] Figure 7 This is an example of a herder management interface of the system configuration interface of the intelligent sheep farming platform in this embodiment of the invention;

[0068] Figure 8 This is a schematic diagram of the breeding performance prediction system based on the intelligent sheep farming platform of the present invention. Detailed Implementation

[0069] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0070] like Figure 8As shown, this embodiment provides a technical solution: a breeding performance prediction system based on a smart sheep breeding platform, including a smart breeding platform (livestock cloud), a data acquisition module, a data processing and analysis module, a machine learning prediction module, a ranking learning optimization module, and a user interaction interface module.

[0071] The above parts will be explained in more detail below.

[0072] 1. Smart farming platform

[0073] like Figure 2 As shown, this platform is developed based on the PHP framework, adopts a B / S architecture, and is deployed on Alibaba Cloud. It includes functional modules such as ranch management, breeding management, data statistics, and traceability system, and supports multi-level user access control (RBAC model).

[0074] In this embodiment, the smart livestock farming platform specifically includes a system configuration module, a basic data management module, a feed formulation management module, a pasture management module, a treatment and disease prevention module, a traceability system module, a breeding management module, a storage management module, a data statistical analysis module, and a data recording module. The system configuration module is the core management module of the platform, used to maintain and manage the basic settings and user permissions of the smart livestock farming platform. The basic data management module is used to maintain and manage the core basic information in the smart livestock farming platform, ensuring the accuracy and consistency of platform data. The feed formulation management module is used to maintain and manage the entire feed formulation process, ensuring the precision and traceability of feed production. The pasture management module is used to comprehensively manage livestock information and related operation records in the pasture, ensuring the accuracy and traceability of livestock data. The treatment and disease prevention module includes modules such as treatment records, disease prevention records, drug files, disease files, and prescription files. The traceability system module is used to comprehensively record and manage the entire process of sheep from acquisition to delivery, ensuring data transparency and traceability at each stage. The breeding management module is used to comprehensively record and manage the breeding process of livestock, ensuring the accuracy and traceability of breeding data. The warehouse management module manages warehouse and inventory information within the ranch, ensuring efficient management and use of supplies. The data statistics and analysis module performs statistical analysis on various data points within the ranch, helping to optimize ranch management. The data recording module comprehensively records all data points within the ranch, ensuring data accuracy and traceability.

[0075] 2. Data Acquisition Module

[0076] like Figures 3-5As shown, the data acquisition module collects data such as sheep body size and weight through RFID ear tags (sheep number (unique identifier), date of birth, sex, breed, introduction time, farm, father / mother ear tag), 3D / 2D cameras (body height, body length, chest circumference, cannon circumference), and dynamic weighing equipment (weight), and stores the data in a standardized manner in a MySQL database.

[0077] Specifically, for rams, the data collected mainly includes growth performance indicators (body weight, height, body length, chest circumference, and cannon bone circumference); for ewes, in addition to the above growth performance indicators, reproductive performance indicators (number of lambs, reproductive survival rate, and twinning rate) are also collected.

[0078] Reproductive performance indicators are collected through the "Lambing Record Module" under the "Breeding Management Module" interface of the smart breeding platform. Specifically, herders or staff enter mating and lambing information into the platform, and the system automatically calculates or records the number of lambs, survival rate, and twinning status.

[0079] 3. Data Processing and Analysis Module

[0080] The data processing and analysis module includes a data preprocessing unit and a statistical analysis unit.

[0081] Data preprocessing unit: Standardizes and complies with data formats, removes outliers, corrects errors, and eliminates duplicate records. Through entity recognition, redundancy elimination, and conflict resolution, it can aggregate data from different sources into a single dataset, facilitating subsequent data analysis and other operations. While preserving as much original data information as possible, it compresses the dataset size, thereby improving the efficiency of analysis and mining, and ensuring that the results obtained are almost identical to those of the original data analysis.

[0082] The statistical analysis unit is used to perform statistical analysis on the individual phenotypic characteristic data of sheep after the data preprocessing unit has cleaned the data. Specifically, Pearson correlation analysis is used to calculate the correlation coefficient between growth and reproductive performance indicators and the Comprehensive Breeding Value (GEBV) to identify the key characteristics that mainly affect the breeding performance of sheep. Furthermore, the path coefficient analysis method is used to quantitatively decompose the direct and indirect effects of each characteristic on GEBV, obtain the causal relationship structure between characteristics, and use it to guide the optimization configuration of characteristic weighting and model input variables, thereby improving the accuracy and stability of the subsequent breeding performance prediction model.

[0083] 4. Machine Learning Prediction Module

[0084] The machine learning prediction module uses machine learning models to predict the overall breeding value of sheep, thereby achieving accurate prediction of breeding performance.

[0085] Machine learning models include ram models and ewe models.

[0086] Ram models: AdaBoost, GBRT, RF, Extra-Trees;

[0087] Models for breeding ewes: SVR, KNN, LR, DTR;

[0088] During system operation, not all models are invoked simultaneously for prediction. Instead, the optimal model is selected based on performance requirements and data characteristics. In the model building phase, the system performs cross-validation and performance evaluation on each candidate algorithm (AdaBoost, GBRT, RF, Extra-Trees, SVR, KNN, LR, DTR, etc.), determining the optimal model by comparing its coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). In the formal prediction phase, only the best-performing model is used to predict the comprehensive breeding value (GEBV). Specifically, if the data characteristics exhibit strong nonlinear relationships and the sample size is large, the GBRT or Extra-Trees model is preferred; if the data dimensionality is low or sensitivity to outliers is high, the RF model is used; if the sample size is relatively limited and the model needs rapid iterative updates, the AdaBoost model is used. The system can automatically select the corresponding model based on historical evaluation results, or researchers can manually set the model type.

[0089] In the ram model, the input data consists of standardized growth performance indicators obtained through the data preprocessing module, including: Body Weight (BW), Body Height (BH), Body Length (BL), Chest Girth (CG), and Tube Bone Radius (TBR). All of these data originate from phenotypic data collected by the smart breeding platform, and are input into the model after missing value repair, outlier correction, and normalization. The output data is the Comprehensive Breeding Value (GEBV) for the corresponding ram individual, representing the individual's genetic breeding potential.

[0090] In the breeding ewe model, the input data consists of preprocessed growth and reproductive performance indicators of the sheep, including growth performance indicators such as body weight (BW), height (BH), body length (BL), and chest circumference (CG), and reproductive performance indicators such as litter size (LS) and survival rate (SR). These data are standardized and feature-selected before being input into each regression model. The output data is the comprehensive breeding value (GEBV) for the corresponding individual breeding ewe.

[0091] During the model performance evaluation phase, the system can further combine the Learning to Rank algorithm with the Subjective-Objective Combined Weighting and Optimization (AHP-PCA) method to perform weighted optimization of the input features. The data reconstructed with these feature weights is then input back into each candidate model to improve the final prediction accuracy. According to the evaluation results, the optimized GBRT model had the highest R² value (approximately 0.94) among the ram models, while the optimized SVR model performed best among the ewe models (R²≈0.78). Therefore, the system is defaulted to using the GBRT model and the SVR model as the final prediction models, respectively.

[0092] 5. Sorting Learning Optimization Module

[0093] A combination of AHP (subjective weighting) and PCA (objective weighting) methods is used to optimize feature weights. Archimedes' optimization algorithm (AOA) is also introduced to adjust the hyperparameters of the machine learning model.

[0094] In this embodiment, the specific process of optimizing feature weights using a combination of AHP (subjective weighting) and PCA (objective weighting) methods is as follows:

[0095] S11: Data preprocessing.

[0096] 1. Standardize the growth performance and reproductive performance indicators of sheep collected by the smart farming platform, including: missing value repair, outlier detection, error correction and duplicate record removal;

[0097] 2. The Z-score standardization method is used to transform each feature into a data form with zero mean and unit variance, resulting in a standardized feature matrix, which provides a unified dimensional basis for subsequent analysis.

[0098] S12: Subjective weighting (AHP weight calculation).

[0099] 1. Based on the experience and knowledge of sheep breeding experts, construct a pairwise comparison matrix A of characteristic indicators, including weight, height, length, chest circumference, cannon bone circumference, number of lambs born, and reproductive survival rate.

[0100] 2. Use the Saaty scaling method (1–9 scale) to score the relative importance of each indicator, solve for the largest eigenvalue and corresponding eigenvector of the matrix, and normalize to obtain the AHP weight vector.

[0101] 3. Calculate the consistency ratio. If CR < 0.10, the matrix is ​​considered consistent; otherwise, return to the previous judgment.

[0102] S13: Objective weighting (PCA weight calculation).

[0103] 1. Perform principal component analysis on the standardized feature matrix to calculate the eigenvalues ​​and loading matrix.

[0104] 2. Select the top m principal components with a cumulative variance contribution rate of not less than 85%, calculate the comprehensive contribution value of each original feature and perform normalization processing to obtain the PCA weight vector.

[0105] S14: Integration of subjective and objective weights.

[0106] 1. Set the subjective-objective fusion coefficient α∈[0,1], with a default value of 0.5. When the expert consensus is good (CR<0.05), it can be increased to 0.7.

[0107] 2. A linear weighting method is used to integrate the weights of AHP and PCA to calculate the combined weight vector, so as to take into account both expert experience and objective data laws.

[0108] S15: Feature Reconstruction.

[0109] Based on the combined weights, the standardized feature matrix is ​​weighted to obtain the weighted reconstructed feature matrix, which is used as the training input for subsequent machine learning models.

[0110] S16: Model training and validation.

[0111] The weighted reconstructed feature matrix and the corresponding Comprehensive Breeding Value (GEBV) were used as input and output samples to train various candidate models: ram models included AdaBoost, GBRT, RF, and Extra-Trees; ewe models included SVR, KNN, LR, and DTR. K-fold cross-validation (K=5 or 10) was used to evaluate the performance metrics (R², RMSE, MAE) of each model, and the best-performing model was selected as the system prediction model.

[0112] S17: Performance evaluation and threshold determination.

[0113] Compare the changes in model performance before and after the AHP-PCA combined weighting optimization. When the R² increase value ΔR² ≥ 0.05 or the RMSE reduction rate ≥ 3%, the optimization is deemed effective and the current weighting scheme is fixed. If the threshold is not reached, return to step S12 or S13 to readjust the weights.

[0114] S18: Online deployment and periodic retraining.

[0115] During operation, the system periodically executes steps S11 to S17 to update the model. When the number of new samples exceeds the set threshold or the time interval exceeds three months, the retraining process is automatically started.

[0116] S19: Exception and Robust Handling.

[0117] 1. When the consistency of the pairwise comparison matrix A fails (CR≥0.10), the system automatically reverts to the PCA single-weighting mode;

[0118] 2. When the cumulative variance contribution rate of PCA is less than 60%, it suggests expanding the sample or introducing new features to improve the information content;

[0119] 3. The system automatically removes invalid features with zero or near-zero variance to prevent weight distortion.

[0120] S110: Result output and storage.

[0121] The final output includes:

[0122] (1) Combine weight vectors;

[0123] (2) Optimal model type and parameters;

[0124] (3) Model performance indicators (R², RMSE, MAE).

[0125] All results are stored in the system database for subsequent breeding performance prediction and model iteration.

[0126] In this embodiment, the specific process of introducing the Archimedes Optimization Algorithm (AOA) to adjust the model hyperparameters is as follows:

[0127] S21: Optimization Problem Definition and Encoding (Input: List of candidate optimization terms; Output: Encoded search space)

[0128] 1. Define the set of variables P to be adjusted in this AOA optimization: In this embodiment, the set of variables P is the set of hyperparameters of the machine learning model;

[0129] 2. Map P to a continuous real vector representation (using integer encoding / mapping functions for discrete parameters if necessary), and define upper and lower bounds for each dimension;

[0130] 3. Define the fitness function F(P): The fitness function is the average RMSE (or average RMSE) obtained from cross-validation on the validation set. The standard is to specify whether regularization terms (such as model complexity penalties) are included.

[0131] S22: Initialize the population and individual physical properties (Input: Population size N, variable dimension Dim, boundary; Output: Initial population and properties D, V, A, where D is density, V is volume, and A is acceleration)

[0132] 1. Randomly generate the position vectors of N candidate solutions (individuals), and uniformly sample each component within its corresponding range;

[0133] The position vector refers to the candidate solution (individual). In breeding prediction, it corresponds to the set of hyperparameters of the model to be optimized (such as learning rate, number of trees, etc.).

[0134] 2. Initialize physical properties for each individual: volume, density, and acceleration (usually randomly initialized within the range of [0,1] or small positive numbers); these properties are used to drive subsequent position updates (refer to the principle of AOA).

[0135] 3. Calculate the fitness of each initial individual and record the current best individual and its fitness.

[0136] S23: Start of main loop — iteration count t←1 (Input: initial population and its physical properties (D, V, A), maximum number of iterations Tmax; Output: update pointer);

[0137] Enter the iteration loop until the stopping condition is met (see S29).

[0138] S24: Update physical properties (density D, volume V, acceleration A) (Input: current individual properties and fitness; Output: updated D, V, A)

[0139] 1. For each individual (particle) k, update the density and volume based on its current fitness and the best fitness of the population—that is, if an individual has a better fitness than the average fitness of the population, its density or volume is adjusted in a favorable direction (the specific implementation can adopt the common AOA update strategy: proportional increase or decrease or scaling based on the difference).

[0140] 2. Calculate the individual acceleration (used to characterize the individual's motion trend), which depends on density, volume, and distance from the optimal individual;

[0141] S25: Determine the search phase (exploration vs. exploitation) and transition operator for the current iteration (Input: current iteration count t, Tmax; Output: transition operator TF, current search phase identifier)

[0142] 1. Set stage probabilities or transition functions based on the iteration progress (e.g., enter the exploration stage with a higher probability in the early stages, and tend to enter the utilization stage in the later stages).

[0143] 2. Use common transfer operators or probability thresholds in AOA to determine whether this is "collision / random movement (Exploration)" or "expitation towards the optimal point".

[0144] S26: Position Update (Input: Current X, D, V, A and stage flag; Output: New position X')

[0145] 1. If in the exploration phase, update the individual position according to the exploration strategy: for example, based on random perturbation or neighborhood sampling, and combine density / volume as weights to adjust the step size to encourage global search;

[0146] 2. If in the utilization phase, update the position towards the optimal individual with acceleration, so that the solution converges in a better direction;

[0147] 3. Perform boundary checks on the updated position and map it back to a valid domain (e.g., truncate to or use a bounce boundary strategy).

[0148] S27: Fitness Evaluation and Population Replacement (Input: X'; Output: Updated Population and Optimal Solution)

[0149] 1. For each new position, de-encode it into actual parameters, build the corresponding model and calculate the fitness in the training / validation process;

[0150] 2. If the position vector is better than the original fitness of the individual, then replace the original individual with the newly generated position vector;

[0151] 3. Update the global optimal individual (if improvements are made); record the convergence curve for subsequent analysis.

[0152] S28: Local Search Enhancement and Fine-tuning (Optional, Input: Current optimal solution; Output: Fine-tuned optimal solution)

[0153] 1. When there is no improvement for several consecutive generations or the preset early convergence criterion is reached, a local search is triggered: fine-grained random perturbation or local gradient search (if differentiable) is applied to the surrounding area to fine-tune the parameters.

[0154] 2. If a better solution is found using local search, then replace the global optimum; and write the improvement into the history archive for retraining / validation.

[0155] S29: Stop condition judgment (Input: t, Tmax, improved threshold; Output: continue / termination flag)

[0156] 1. When the fitness of the globally optimal individual improves less than a threshold over consecutive generations. Then terminate the iteration; otherwise, let t←t+1 and return to S24;

[0157] 2. Upon termination, output the final optimal parameters, corresponding model performance indicators, and convergence history curves.

[0158] S210: Result Validation, Deployment, and Rollback Strategies (Inputs: P^{*} and training logs; Outputs: Deployment plan and risk control)

[0159] 1. Final validation will be performed on an independent test set, and metrics such as R², RMSE, and MAE will be recorded. If the preset performance targets are met (e.g., ΔR² ≥ 0.05 or RMSE decreases by ≥ 3%), the model will be written into the model library and deployed as online prediction parameters.

[0160] 2. If the objective is not achieved, alternative strategies can be initiated: use the AHP-PCA combined weighting results as a baseline (see S11–S20), or expand the search space, increase the population size and number of iterations, and then repeat S21–S29;

[0161] 3. Record all experimental configurations, population logs, and random seeds to ensure the experiment is reproducible.

[0162] 6. User Interface Module

[0163] The charts and graphs display sheep growth trends and breeding performance predictions, and the interface is designed to support multiple roles such as herders, managers, and researchers.

[0164] In this embodiment, the data flow transmission process of this system is as follows: data acquisition module → MySQL database → data processing and analysis module → machine learning model → ranking learning optimization → user interface display.

[0165] In this embodiment, the front-end (HTML5 / CSS3 / JavaScript) interacts with the database through a PHP back-end, and the machine learning model (Python) processes the data and returns the prediction results.

[0166] In this embodiment, the sheep's body size and RFID reader communicate in real time with the breeding performance prediction system via Internet of Things (IoT) technology.

[0167] In this embodiment, the specific operation process is as follows:

[0168] (1) Data entry: Herdsmen manually enter or automatically collect sheep body size and breeding data through the platform.

[0169] (2) Data analysis: The system automatically preprocesses the data and generates statistical analysis reports (such as correlation heatmaps).

[0170] (3) Model prediction: The user selects a prediction model (such as the GBRT ram model), inputs data, and obtains the GEBV prediction value.

[0171] (4) Optimization and adjustment: Enable AHP-PCA combined weighting or AOA optimization, and retrain the model to improve accuracy.

[0172] (5) Application of results: Select high-quality breeding sheep based on the prediction results, formulate breeding plans, and visualize the data.

[0173] The breeding performance prediction system of this invention can significantly improve breeding management efficiency, economic benefits, and breeding prediction accuracy, and has the advantages of significant social benefits and strong technical extensibility, as detailed below:

[0174] Data collection efficiency is improved by over 80%. Traditionally, manually recording the body size data of each sheep takes 5-10 minutes. This system, through IoT devices (such as a 3D body size measuring instrument), can automatically collect data within 30 seconds, and the data is automatically synchronized to the cloud. The RFID ear tag recognition speed is 0.5 seconds per ear, which is 10 times faster than traditional ear tag recognition. In addition, management costs are reduced by 60%, reducing the need for dedicated record-keeping personnel.

[0175] Direct economic benefits are improved: through accurate forecasting, the utilization efficiency of breeding rams increases by 30%, reducing losses from ineffective mating and improving the overall utilization rate of breeding rams; adjusting feeding programs based on forecast results reduces the feed conversion ratio from 4.5:1 to 4.0:1, saving approximately 500,000 yuan in feed costs annually for farms with 10,000 head of sheep, and optimizing feed conversion rates. Indirect economic benefits include: the selling price of certified high-quality breeding rams increases by 20-30%, resulting in a premium for high-quality breeding rams; the rate of genetic progress increases by 40%, shortening the breeding cycle by 2-3 years and accelerating herd improvement.

[0176] The actual selection accuracy has been improved. In the comparative test at the Sunite Left Banner demonstration base, the identification rate of high-quality breeding sheep selected by traditional experience was only 65%, while the identification rate of high-quality breeding sheep predicted by this system reached 92%, and the average daily weight gain of offspring sheep increased by 17%.

[0177] Improve environmental benefits; precise feeding reduces methane emissions by 5%; electronic management reduces paper record-keeping by 3 tons per 10,000 head of farms annually.

[0178] Multi-species compatibility: Successfully adapted to breeds such as cashmere goats and Hu sheep, with model migration costs below 10%. Functional scalability: Reserved interfaces for disease prediction, feed formulation optimization, etc., allowing for rapid expansion of new modules.

[0179] In summary, the breeding performance prediction system based on the intelligent sheep farming platform described in the above embodiments firstly developed the "Livestock Cloud" intelligent farming platform, realizing standardized collection, storage, and intelligent management of sheep life-cycle data, laying a solid data foundation for subsequent research. Secondly, it innovatively combines ensemble learning (AdaBoost, GBRT, etc.), machine learning (SVR, KNN, etc.), and ranking learning algorithms to construct a multi-level sheep breeding performance prediction model system. Specifically, it proposes a subjective and objective combined weighting method based on AHP-PCA, significantly improving the model's prediction accuracy through feature reconstruction. Furthermore, it introduces the Archimedes Optimization Algorithm (AOA) into the field of breeding prediction for the first time, utilizing its physical simulation mechanism to achieve adaptive hyperparameter search, significantly improving the model's accuracy, efficiency, and robustness. These innovations not only fill the gap in the application of ranking learning in animal husbandry but also provide a new solution for achieving intelligent, data-driven precision breeding. The "Livestock Cloud" intelligent farming platform was designed to achieve standardized collection and management of sheep life-cycle data.

[0180] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A breeding performance prediction system based on a smart sheep farming platform, characterized in that, It includes a smart farming platform, a data acquisition module, a data processing and analysis module, a machine learning prediction module, a ranking learning optimization module, and a user interface module. The smart farming platform is developed based on the PHP framework, adopts a B / S architecture design, is deployed on the Alibaba Cloud platform, and supports multi-level user access control. The data acquisition module is used to collect sheep-related data and store the data in a standardized manner in a MySQL database; the data processing and analysis module is used to perform data preprocessing and data analysis. The machine learning prediction module is used to predict the comprehensive breeding value of sheep through machine learning models, and obtain the breeding performance prediction results of sheep; the ranking learning optimization module is used to retrain the machine learning model to improve accuracy by using AHP-PCA combined weighting or AOA optimization methods; the user interface module is used to display the growth trend of sheep and the breeding performance prediction results through charts.

2. The breeding performance prediction system based on a smart sheep farming platform according to claim 1, characterized in that, The smart farming platform includes a system configuration module, a basic data management module, a feed formulation management module, a pasture management module, a treatment and disease prevention module, a traceability system module, a breeding management module, a storage management module, a data statistical analysis module, and a data recording module. The system configuration module is used to maintain and manage the basic settings and user permissions of the smart farming platform; the basic data management module is used to maintain and manage the core basic information in the smart farming platform. The feed formulation management module is used to maintain and manage the entire feed formulation process; the ranch management module is used to comprehensively manage livestock information and related operation records in the ranch; the treatment and disease prevention module includes treatment records, disease prevention records, drug files, disease files, and prescription files; the traceability system module is used to comprehensively record and manage the entire process of sheep from acquisition to delivery; the breeding management module is used to comprehensively record and manage the breeding process of livestock; and the warehousing management module is used to manage warehouse and inventory information in the ranch. The data statistics and analysis module is used to statistically analyze various data in the ranch; the data recording module is used to comprehensively record various data in the ranch.

3. The breeding performance prediction system based on a smart sheep farming platform according to claim 2, characterized in that, In the data acquisition module, for rams, the collected data includes growth performance indicators, such as body weight, height, length, chest circumference, and cannon bone circumference; for ewes, in addition to the growth performance indicators, reproductive performance indicators are also collected, such as number of lambs, survival rate, and twinning rate.

4. The breeding performance prediction system based on a smart sheep farming platform according to claim 3, characterized in that, The data processing and analysis module includes a data preprocessing unit and a statistical analysis unit. The data preprocessing unit standardizes and complies with data formats, removes outliers, corrects errors, and eliminates duplicate records. Through entity recognition, redundancy elimination, and conflict resolution, it aggregates data from different sources into a single dataset and compresses the dataset. The statistical analysis unit performs statistical analysis on the individual sheep phenotypic characteristic data cleaned by the data preprocessing unit. Specifically, Pearson correlation analysis is used to calculate the correlation coefficients between growth performance indicators, reproductive performance indicators, and the overall breeding value to identify the main characteristics affecting sheep breeding performance. Then, path coefficient analysis is used to quantitatively decompose the direct and indirect effects of each characteristic on the overall breeding value, obtaining the causal relationship structure between characteristics. This structure guides the optimization of feature weighting and model input variables, thereby improving the accuracy and stability of the subsequent breeding performance prediction model.

5. The breeding performance prediction system based on a smart sheep farming platform according to claim 4, characterized in that, In the machine learning prediction module, the machine learning models include a ram model and a ewe model; Ram models include AdaBoost, GBRT, RF, and Extra-Trees; During runtime, the optimal model is selected from AdaBoost, GBRT, RF, and Extra-Trees based on performance requirements and data characteristics. In the model building phase, cross-validation and performance evaluation are performed on each candidate ram model. The optimal model is determined by comparing their coefficient of determination R², root mean square error RMSE, and mean absolute error MAE. During the formal prediction phase, the optimal model is invoked to predict the comprehensive breeding value. The model input is the growth performance index of the breeding ram, and the output data is the comprehensive breeding value of the corresponding individual breeding ram, which is used to represent the genetic breeding potential of the corresponding individual. Breeding ewe models include SVR, KNN, LR, and DTR; During runtime, the optimal model is selected from SVR, KNN, LR, and DTR based on performance requirements and data characteristics. In the model building phase, cross-validation and performance evaluation are performed on each candidate ewe model. The optimal model is determined by comparing its coefficient of determination R², root mean square error RMSE, and mean absolute error MAE. During the formal prediction phase, the optimal model is invoked to predict the comprehensive breeding value. The model input consists of the growth performance index and reproductive performance index of the breeding ewe, and the output data is the comprehensive breeding value of the corresponding individual breeding ewe, which is used to represent the genetic breeding potential of the corresponding individual.

6. The breeding performance prediction system based on a smart sheep farming platform according to claim 1, characterized in that, In the ranking learning optimization module, the specific process of optimization using AHP-PCA combined weighting is as follows: S11: Data Preprocessing The collected growth performance indicators and reproductive performance indicators were standardized to obtain a standardized feature matrix; S12: Subjective Empowerment Based on the experience and knowledge of sheep breeding experts, pairwise comparison matrices A for characteristic indicators were constructed for breeding rams and breeding ewes respectively. For breeding rams, the indicators included weight, height, length, chest circumference, and cannon bone circumference; for breeding ewes, the indicators included weight, height, length, chest circumference, cannon bone circumference, number of lambs born, reproductive survival rate, and twinning rate. The relative importance of each indicator was scored using the Saaty scaling method, the largest eigenvalue and corresponding eigenvector of the matrix were solved, and the matrix was normalized to obtain the AHP weight vector. The consistency ratio CR was calculated. When CR < 0.10, the matrix was considered to be consistent; otherwise, the evaluation was returned to the next step. S13: Objective Empowerment Principal component analysis is performed on the standardized feature matrix to calculate the eigenvalues ​​and loading matrix. The top m principal components with a cumulative variance contribution rate of not less than 85% are selected, and the comprehensive contribution value of each original feature is calculated and normalized to obtain the PCA weight vector. S14: Integration of Subjective and Objective Weights Set the subjective and objective fusion coefficient α∈[0,1], and use the linear weighting method to fuse the AHP and PCA weight vectors based on the subjective and objective fusion coefficients to calculate the combined weight vector; S15: Feature Reconstruction Based on the combined weight vector, the standardized feature matrix is ​​weighted to calculate the weighted reconstructed feature matrix; S16: Model Training and Validation The weighted reconstructed feature matrix and the corresponding comprehensive breeding value were used as input and output samples to train various ram and ewe models. K-fold cross-validation was used to evaluate the performance of each model and select the optimal model. S17: Performance Evaluation and Threshold Determination Compare the changes in model performance before and after AHP-PCA combined weighting optimization. When the R² increase value ΔR² ≥ 0.05 or the RMSE reduction rate ≥ 3%, the optimization is deemed effective and the current weighting scheme is fixed. If the threshold is not reached, return to step S12 or S13 to readjust the weights; S18: Online Deployment and Periodic Retraining During operation, steps S11 to S17 are executed periodically to update the model. When the number of new samples exceeds the set threshold or the time interval exceeds three months, the retraining process is automatically started. S19: Exception and Robust Handling When the consistency of the pairwise comparison matrix A fails, revert to the PCA single weighting mode; when the cumulative variance contribution rate of PCA is less than 60%, suggest expanding the sample or introducing new features to improve the information content; remove invalid features with zero or near-zero variance. S110: Result Output and Storage The output results, including the combined weight vector, the optimal model type and parameters, and the model performance indicators, are stored in the database for subsequent breeding performance prediction and model iteration.

7. The breeding performance prediction system based on a smart sheep farming platform according to claim 1, characterized in that, The specific process of AOA optimization in the ranking learning optimization module is as follows: S21: Definition and Coding of the Optimization Problem Determine the set of variables P to be adjusted in this AOA optimization, which includes the set of hyperparameters of the machine learning model; map P to a continuous real number vector representation, define the upper and lower bounds for each dimension, and finally define the fitness function; S22: Initialize population and individual physical attributes Randomly generate N candidate solution position vectors, and uniformly sample each position vector within the corresponding range. The candidate solution is the individual. Initialize physical properties for each individual to drive subsequent position updates. The physical properties include density, volume and acceleration. Calculate the fitness of each initial individual and record the current best individual and its fitness. S23: Start the main loop and begin iteration. Enter the iterative loop until the stopping condition is met; S24: Update physical properties For each individual k, update the density and volume based on its current fitness and the population's best fitness: if the individual has a better fitness than the population average, adjust its density or volume in a favorable direction; then calculate the individual acceleration. S25: Determine the search phase of the current iteration. Set the stage probability or transition function according to the iteration progress, and use the transition operator or probability threshold in AOA to determine whether this is "collision / random movement" or "acceleration towards the optimal state"; S26: Location Update If in the exploration phase, update the individual position according to the exploration strategy; if in the utilization phase, update the position towards the optimal individual according to the acceleration. Perform boundary checks on the updated position and map it back to the legal domain. S27: Fitness Assessment and Population Replacement For each new position, the actual parameters are decoded. In the training / validation process, the corresponding model is built and the fitness is calculated. If the fitness is better than the original fitness of the individual, the original individual is replaced and the global optimal individual is updated. The global optimal individual is the global optimal solution. S28: Local Search Enhancement and Fine-tuning When there is no improvement for several consecutive generations or the preset early convergence criterion is reached, a local search is triggered. If a better solution is obtained by using the local search, the global optimal solution is replaced. S29: Stopping Condition Judgment When the fitness of the globally optimal individual improves less than a threshold over consecutive generations. If the iteration is terminated, otherwise return to step S24 to continue the iteration. When the iteration is terminated, the final optimal parameters, the corresponding model performance index, and the convergence history curve are output. S210: Result Validation, Deployment, and Rollback Strategies Perform final validation on an independent test set and record the R², RMSE, and MAE metrics. If the preset performance target is met, it will be written into the model library and deployed as an online prediction parameter; If the preset performance target is not met, an alternative strategy is initiated: use the AHP-PCA combined weighting result as a baseline, or expand the search space, increase the population size and number of iterations, and then repeat steps S21 to S29.

Images