基于传感器监测数据的预测泌乳牛乳房炎机器学习算法研究

Study on machine learning algorithm for predicting lactating cow mastitis based on sensor monitoring data

为了预测集约化牧场奶牛患有临床乳房炎的风险,试验选择黑龙江省黑河市两个集约化奶牛场的288头泌乳牛及其数据为研究对象,分为患病组(确诊患有乳房炎的奶牛189头)和健康组(109头),选择这些奶牛的日平均产奶量、日平均活动量、日平均反刍时间、日白天平均反刍时间、日夜间平均反刍时间、昼夜反刍时间比、日每2 h的反刍时间偏差绝对值、日加权反刍时间变化绝对值的和、日平均电导率变化百分比、日电导率峰值14个指标数据,比较分析上述变量在患病组和健康组组间和组内差异;然后采用3种机器学习算法(决策树、随机森林、eXtreme Gradient boosting)和二元逻辑分类算法预测奶牛乳房炎的发病情况。结果表明:在d-0时,患病组奶牛的日平均产奶量[(34.89±11.81)kg]极显著低于健康组[(41.96±8.69)kg,P<0.01];而在d-7~d-2时,患病组奶牛的日平均产奶量均高于健康组,但差异不显著(P>0.05)。患病组奶牛的日平均反刍时间、日白天平均反刍时间、日夜间平均反刍时间在d-1时均达到最低[(515.37±66.88)min、(206.63±67.05)min、(309.56±64.52)min],分别比健康组奶牛[(560.68±51.30)min、(225.81±34.04)min、(334.38±39.89)min]平均少45.31 min(P<0.05)、19.18 min(P>0.05)、24.82 min(P<0.05)。患病组奶牛的昼夜反刍时间比、日每2 h反刍时间偏差绝对值、日加权反刍时间变化绝对值的和均在d-0时达到最大值,而健康组奶牛的上述3个指标从d-7~d-0均无明显变化。患病组奶牛的日加权反刍时间变化绝对值的和在d-1和d-0时分别比健康组极显著高48.83和94.27(P<0.01)。患病组奶牛的日电导率变化百分比、日电导率峰值从d-3~d-0逐渐增大,d-0时达到最大值,而健康组奶牛上述指标均无明显变化。随机森林模型的Se值最高,二元逻辑分类模型最低;eXtreme Gradient boosting模型Sp值高于随机森林模型,但随机森林模型的Acc、F1和AUC值优于其他3种模型。说明随机森林算法对奶牛乳房炎的预测效果最优,日平均产奶量、昼夜反刍比、日加权反刍时间变化绝对值的和、日电导率变化百分比、日电导率峰值可作为奶牛乳房炎预测因子。

In order to predict the risk of intensive grazing cows with clinical mastitis,in this experiment,288 lactating cows from two intensive dairy farms in Heihe City,Heilongjiang Province were selected as research objects and divided into the diseased group(189 cows diagnosed with mastitis)and the healthy group(109 cows).The 14 index data of these cows(the daily average milk production,daily average activity level,daily average rumination time,daily average rumination time at daytime,daily average rumination time at nighttime,day-night rumination time ratio,absolute value of rumination deviation every 2 h per day,sum of absolute value of daily weighted rumination time change,daily average conductivity change percentage,and daily high conductivity peak),were selected to compare and analyze the differences between the groups.Three machine learning algorithms(decision trees,random forest,and eXtreme Gradient boosting)and binary logical classification algorithms were then used to predict the onset of mastitis in dairy cows.The results showed that at the d-0 time,the daily average milk production of the cows in the diseased group([34.89±11.81]kg)was significantly lower than that in the healthy group([41.96±8.69]kg,P<0.01),while at the time of d-7 to d-2,the daily average milk production of the cows in the diseased group was higher than that of the healthy group,but the difference was not significant(P>0.05).The daily average rumination time,daily average rumination time at day time,and daily average rumination time at nighttime([515.37±66.88]min,[206.63±67.05]min,[309.56±64.52]min)in the diseased group reached the lowest at d-1,which were lower than those in the healthy group([560.68±51.30]min,[225.81±34.04]min,[334.38±39.89]min],respectively.The average shortening values were 45.31 min(P<0.05),19.18 min(P>0.05),24.82 min(P<0.05),respectively.The day-night rumination time ratio,the absolute value of rumination deviation every 2 h per day,and the sum of absolute value of daily weighted rumination time change of the diseased group cows all reached the maximum value at d-0,while the above three indicators of the healthy group cows did not change significantly from d-7 to d-0.The sum of absolute value of daily weighted rumination time change of the diseased group cows was significantly higher than that in the healthy group cows(48.83 and 94.27)at d-1 and d-0(P<0.01),respectively.The daily average conductivity change percentage and the daily high conductivity peak of the dairy cows in the diseased group gradually increased from d-3 to d-0,and reached the maximum value at the time of d-0,while the above indicators of the dairy cows in the healthy group did not change significantly.The random forest model had the highest Se value and the binary logical classification model was the lowest.The Sp value of the eXtreme Gradient boosting model was higher than that of the random forest model,but the Acc,F1 and AUC values of the random forest model were better than the other 3 models.The results indicated that the random forest algorithm has the best prediction effect,and the daily average milk production,the day-night rumination time ratio,the sum of absolute values of the daily weighted ruminant time change,the daily average conductivity change percentage,and the daily high conductivity peak could be used as predictors of mastitis in dairy cows.