基于加速度计的肉牛次级采食行为自动识别方法研究

Research on automatic recognition method of beef cattle’s gradational feeding behavior based on accelerometer

关键词:  加速度器  采食行为  XGBoost  HMM

Key words:  accelerator  feeding behavior  XGBoost  HMM

摘要:

【目的】 肉牛采食行为包括卷食、咀嚼、卷食—咀嚼等几种次级行为。监测肉牛次级采食行为有助于评估牛只的健康状况和营养水平。文章旨在利用加速度传感器研究肉牛次级采食行为识别方法，对比不同监测部位对次级采食行为识别的影响。【方法】 将加速传感器安装在肉牛的鼻子、右颌、左嘴3个部位，检测次级采食行为的加速度信号，经过衍生变量函数计算，扩充数据维度，使用ExtraTreesClassifer选择出9种重要特征，运用XGBoost算法识别肉牛采食次级行为（卷食、咀嚼、卷食—咀嚼、其他），最后使用HMM-viterbi算法修正次级行为识别结果。【结果】 XGBoost和HMM-viterbi在鼻子、右颌、左嘴3个部位识别的平均结果相同，XGBoost识别的平均准确率、精确率、F1得分和召回率分别为0.95、0.93、0.93和0.93，HMM-viterbi修正后识别的平均准确率、精确率、F1得分和召回率均为0.99。因此，运用HMM-viterbi模型可以有效修正行为识别结果。在XGBoost识别结果中，鼻子部位识别次级行为的得分较高，考虑长期佩戴传感器的稳定性，推荐采用鼻子作为检测部位。【结论】 在肉牛鼻子部位佩戴加速度器，利用XGBoost结合HMM-viterbi的方法可以自动识别肉牛次级采食行为。

Abstract:

Purpose The feeding behavior of beef cattle includes several gradational behaviors：rolling，chewing，and rolling-chewing. Monitoring the gradational feeding behavior of beef cattle is helpful to assess the health and nutritional level of the beef cattle. This article aims to use acceleration sensors to study the recognition method of gradational feeding behavior of beef cattle，and to compare the influence of different monitoring positions on the recognition of gradational feeding behavior.Method In the experiment，the acceleration sensor was installed on the nose，right jaw，and left mouth of the beef cattle to detect the acceleration signal of the gradational feeding behavior. After calculating the derivative variable function，the data dimension was expanded，and 9 types were selected using ExtraTreesClassifer For important features，the XGBoost algorithm is used to predict the gradational behavior of beef cattle （rolling，chewing，rolling-chewing，other），and finally the HMM-viterbi algorithm is used to modify the gradational behavior prediction results.Result XGBoost and HMM-viterbi predict the same average results in the nose，right jaw，and left mouth. The average accuracy，precision，F1 score and recall rate of XGBoost prediction are 0.95，0.93，0.93 and 0.93，respectively. The average accuracy，precision，F1 score and recall rate of HMM-viterbi's revised forecast are all 0.99. Therefore，the use of the probability directed graph model can effectively modify the behavior prediction results. In the XGBoost prediction results，the nose part has a higher score for predicting gradational behaviors. Considering the stability of wearing the sensor for a long time，it is recommended to use the nose as the detection part.Conclusion Wearing an accelerometer on the nose of beef cattle，using XGBoost combined with HMM-viterbi method can automatically identify the gradational feeding behavior of beef cattle.