Machine learning to detect behavioural anomalies in dairy cows under subacute ruminal acidosis
Machine learning to detect behavioural anomalies in dairy cows under subacute ruminal acidosis
复制标题
机器学习检测亚急性瘤胃酸中毒下奶牛的行为异常
DOI：
10.1016/j.compag.2020.105233
复制
发表时间：
2020
期刊：
影响因子：
--
通讯作者：
I. Veissier
中科院分区：
文献类型：
--
作者：
N. Wagner;Violaine Antoine;M. Mialon;R. Lardy;M. Silberberg;J. Koko;I. Veissier
疾病行为的特征是昏睡状态，在此期间动物减少活动，睡得更多，有时正常清醒，减少饲料和水的摄入量，并减少与环境的相互作用。行为上的细微变化可以在疾病的临床症状出现之前实现。最近传感器的发展使动物行为的持续监测成为可能，但异常动物活动的转变仍然难以检测。我们探索了使用机器学习（ML）从连续监控中检测异常行为。我们提交了14头奶牛（牛）亚急性瘤胃酸中毒（SARA），一种已知会引起行为变化的疾病。另外14头对照奶牛未提交给SARA。我们使用瘤胃推注来监测pH值，并检测奶牛何时出现SARA。我们使用了一个定位系统来推断动物的活动的基础上，它的位置与特定的元素在谷仓（饲料，休息区，和小巷）。我们测试了几种ML算法：K最近邻回归（KNNR）;决策树回归（DTR）;多层感知器（MLP）;长短期记忆（LSTM）;以及一种假设活动从一天到下一天相似的算法。首先，我们开发了ML模型来预测前24小时给定一天的活动，同时考虑所有奶牛。然后，我们计算了给定奶牛的观测值和预测值之间的误差。最后，我们将误差与选择的阈值进行比较，以优化正常值和异常值之间的区别。KNNR表现最好，检测到83%的SARA病例（真阳性），但也产生了66%的假阳性，这限制了它在实践中的使用。总之，ML可以帮助检测行为中的异常。通过在动物而不是群体水平的非常大的数据集上应用ML，可能会获得进一步的改进。
Sickness behaviour is characterised by a lethargic state during which the animal reduces its activity, sleeps more and at times when normally awake, reduces its feed and water intake, and interacts less with its environment. Subtle modifications in behaviour can materialise just before clinical signs of a disease. Recent sensor developments enable continuous monitoring of animal behaviour, but the shift to abnormal animal activity remains difficult to detect. We explored the use of Machine Learning (ML) to detect abnormal behaviour from continuous monitoring. We submitted 14 cows (Bos taurus) to Sub-Acute Ruminal Acidosis (SARA), a disease known to induce changes in behaviour. Another 14 control cows were not submitted to SARA. We used a ruminal bolus to monitor pH and detect when a cow experienced SARA. We used a positioning system to infer an animal’s activity based on its position in relation to specific elements in the barn (feeder, resting area, and alleys). We tested several ML algorithms: K Nearest Neighbours for Regression (KNNR); Decision Tree for Regression (DTR); MultiLayer Perceptron (MLP); Long Short-Term Memory (LSTM); and an algorithm where activity is assumed to be similar from one day to the next. First, we developed ML models to predict activity on a given day from the previous 24 h, considering all cows together. Then, we calculated the error between observed and predicted values for a given cow. Finally, we compared the error to a threshold chosen to optimise the distinction between normal and abnormal values. KNNR performed best, detecting 83% of SARA cases (true-positives), but it also produced 66% of false-positives, which limits its use in practice. In conclusion, ML can help detect anomalies in behaviour. Further improvements could probably be obtained by applying ML on very large datasets at animal rather than group level.
DOI：
--
发表时间：
2007
期刊：
--
影响因子：
--
作者：
R. Dantzer;K. Kelley
通讯作者：
R. Dantzer;K. Kelley