基于姿态角的生猪行为识别方法研究

Research on pig's behavior recognition based on attitude angle

现有的生猪行为判别方法,未能实现个体全天行为的监测统计,且模型分析过程繁琐、准确率低、设备成本高。研究低成本、高准确率的生猪个体行为识别方法,为判别生猪健康状况提供依据。首先设计基于微惯性传感器且可穿戴生猪姿态信息监测模块,将监测模块佩戴在生猪颈部,采集站、走、卧、躺4类生猪行为下对应的加速度、角速度、姿态角数据6 000组,然后采用LM训练法进行模型训练,并分析加入姿态角信息对模型训练结果的影响,最后建立生猪日常行为分类模型并采用异校验方式进行验证。结果表明:在考虑姿态角信息的条件下,其训练效果与模型的相关系数均优于仅考虑行为信息的训练模型,加入姿态角信息作为BP神经网络输入量可有效避免网络陷入到局部极小值,收敛迅速,误差函数为0.001 844,满足训练要求,模型期望分类与实测分类的决定系数为0.991,生猪行为识别准确率为92.64%。该模型为进一步分析生猪生理健康状况判别提供了数据支持。

The existing way of pig＇s behavior recognition could not realize pig＇s activities monitor throughout the day,and some problems such as complex model,low accuracy and high-costs in recognition method remains unsolved.Therefore,we designed low cost and high accuracy pig＇s behavior recognition,which provided the basis on identifying the pig＇s health.Firstly,we designed wearable pig＇s gesture information monitoring module based on MEMS,which was worn on the neck of pig,collecting 6 000 experimental data of acceleration,angular and attitude angle with pig＇s standing,walking,side lying and lying.Then,Levenberg-Marquardt（LM）training method was used.Meanwhile,the effect of attitude angle on the training results was analyzed.Finally,different calibrations were used to validate the pig daily behaviors classification model.The result showed that the training results of the training results of the training method considering attitude angle information was superior to the training model,which was only considered the pig＇s behaviors information.The model can effectively avoid network into a local minimum value,considering attitude angle information as input of BPNN.The error rate was 0.001 844 and the determination coefficient between actual measured and calculated values was 0.991 and the behavior recognition accuracy was 92.64%.The result indicated that pig＇s daily behavior recognition based on attitude angle can provide a theoretical basis for analyzing pig＇s physical health.