- 期刊
應用機械學習技術預測母豬第三胎出生活仔數
Application of Machine Learning Technologies to Predict Sows' Number of Piglets Born Alive (NBA) at 3rd Parity
摘要
正確評估母豬繁殖性能為維持豬場生產效能之重要工作，惟飼養管理者常僅憑藉自身主觀經驗及有限之簿記資料，進行母豬選留；此方法易造成錯估後續胎次繁殖性能，增加母豬淘汰風險，蒙受經濟損失。本研究應用機械學習技術與場內簿記資料，建立母豬繁殖性能預測模型，早期預測母豬第三胎出生活仔數（Number of piglets born alive；NBA）表現。母豬資訊蒐集自我國一商業豬場，涵蓋797隻母豬自出生後之歷史資料，共計36種變數；其中，第三胎出生活仔數分為高組（畜群中最佳的前33%）與中低組（剩餘67%），作為擬預測之繁殖性能。應用四種機械學習演算法建立模型，包括：（1）線性判別分析（Linear Discrimination Analysis, LDA）、（2）邏輯斯回歸（Logistic Regression, LR）、（3）類神經網路（Neural Network, NN）及（4）梯度提升樹（Gradient Boost tree, GB），搭配五折交叉驗證（five-fold cross validation）評估模型預測性能。結果顯示，梯度提升樹模型預測效能最佳，準確度達78%。後應用permutation importance方法分析各變數之重要性，前五項影響預測效能最鉅之變數依序為：（1）第一胎出生活仔數、（2）第二胎離乳前死亡率、（3）第一胎懷孕天數、（4）第二胎離乳至懷孕天數及（5）初胎配種懷孕日齡。機械學習技術可就母豬早期表現，預測其第三胎繁殖性能，作為選留母豬之決策輔助工具，協助提升牧場豬群繁殖性能。
並列摘要
Accurate prediction of the reproductive performance of sows is crucial to a pig farm's benefit. In practice, the stock person often relies on limited records and personal experience to carry out sow selection for breeding. Such a method is biased and may restrict the selection accuracy, potentially affecting the benefit. This study applied machine learning techniques and sow records to predict the reproductive performance of sows. The target of the prediction was the number of piglets born alive (NBA) of 3rd parity sows. In total, 36 performance variables derived from 797 sows were collected to build the machine learning model. The NBA was classified into high- (the ranking of the top 33%) and low-groups (the rest 67%) as the prediction target. Four machine learning methods were utilized, including Linear Discrimination Analysis (LDA), Logistic Regression (LR), Neural Network (NN), and Gradient Boost tree (GB). According to the results of five-fold cross validation, GB presents the best prediction performance compared with other methods, and the prediction accuracy was 78%. Given the results from permutation importance analysis, NBA at 1st parity, Pre-weaning mortality (PWM) at 2nd parity, pregnancy length at 1st parity, Wean-to-Mate Interval (WMI) at 1st parity, and age of first successful service were the top five important variables. In conclusion, the application of machine learning technologies and reproductive performance records can build predictive models to support the stock person in decision-making and to improve farm productivity.