基于声信号的母羊行为特征识别及数字化表达方法研究

Inner Mongolia and its surrounding west area are the main districts of sheep production. However, the traditional grazing manner caused the problems of grassland degradation, waste of forage resources as well as low breeding efficiency, so the development of intensive and large-scale sheep farming is the present tendency. Due to high density of sheeps in farming house, in case of emergency such as ewe delivery, it will cause a great loss if breeder dosen't know its happening and takes measures in time.Considering that sheep will make different sound in different emergent situations, the sound can not only reflect the status of sheep's organism or health but also can reflect the response behaviour to the environmental sudden change. This proposal will establish an ewe's sound signal acquisition system based on wireless sensor network, and build a relationship between digital feature of ewe's sound and its emergency response behaviors through voice recognition and ewe's sound analysis, feature extraction and digital expression, in case of emergency response behaviors, such as ewe delivery, ewe rumination, ewe looking for baby, illness, hungery or thirst and ewe suffered attack (Captive sheep often suffer attack by wild animals and theft in the rural and pastoral areas), the results will provide a theoretical basis for developing the emergency response system of the ewe breeding.

内蒙古及周边西部地区是我国养羊业的主产区，但传统的放牧饲养及农户散养方式造成了草场退化、资源浪费、养殖效益低的问题，因此发展集约化、规模化种草圈养，是目前养羊业的发展趋势。但规模化圈养羊的密度大，在应急情况（如集中产羔期）饲养员不能及时发现并采取措施，会造成很大损失。因羊在不同的应激行为情况下，会发出不同的叫声或声音，其声信息能反映其机体的状况及其对环境的应激反应。本项目通过建立基于无线传感网络的母羊发声声音信号采集系统，对母羊在不同行为情况下，如母羊分娩或难产、母羊反刍、生病、饥饿或饥渴、以及母羊受到侵袭（在农牧区圈养羊常常遭受偷盗和野生动物的袭击）等，发出的声音信号进行时域和频域分析、特征提取与数字化表达，建立母羊声音信号数字化特征与母羊不同行为的相关性，为母羊福利化圈养养殖和应急预警系统的研究提供理论依据。

内蒙古地区传统的放牧饲养及农户散养方式造成了草场退化和养殖效益低的问题，集约化设施养羊是养羊业的发展趋势。在规模化设施圈养条件下，由于羊的饲养密度大、活动范围受限和环境调控不到位等因素，羊易产生应激行为和健康问题，羊声信号包含了羊只对其内部机体状况和需求的信息反馈。项目通过无线声音数据采集平台，采集设施母羊舍内的打斗、饥饿、咳嗽、啃咬、反刍、寻羔和寻伴等声信号，并进行声信号的小波阈值去噪、特征参数提取和分类识别研究主要研究，内容和结果如下：(1)对母羊舍中风机、饲喂设备等产生多大噪声，采用小波阈值去噪，并对小波的阈值选择规则和阈值函数进行改进，当母羊声信号受风机噪声干扰时，改进的小波阈值去噪算法取得了较好的去噪性能。(2)针对用传统线性预测系数算法提取母羊声信号共振峰时，存在合并峰和虚假峰的问题，项目将母羊的声道看成是若干个谐振腔串联，用改进的线性预测系数算法提取声信号共振峰，通过对共振峰轨迹跟踪曲线的分析，得到了母羊声信号特征参数。(3)对母羊声信号的梅尔频率倒谱系数及其一阶差分构成的混合参数进行改进，提出基于特征相关性的Fisher准则与特征分量加权相结合的混合特征参数，该混合特征参数在低维度下也能取得较高的识别率。(4)将希尔伯特‒黄变换应用到母羊声信号的特征参数提取中，对其进行进行经验模态分解得到7个固有模态函数分量，对每个固有模态函数分量求其边际谱能量倒谱系数，再结合Mel尺度曲线进行频域划分，得到的子带能量倒谱系数作为特征参数，可提高母羊声信号的识别率。(5)利用隐马尔科夫模型的动态时序建模能力和BP神经网络的分类决策能力，构建隐马尔科夫模型/BP神经网络的混合识别模型，将隐马尔科夫模型的输出状态累积概率作为BP神经网络的输入，得到了比单一识别模型更高的识别率。研究的成果结合开发的基于物联网技术设施养殖羊舍环境调控系统，将声信号应用于监测和评价其养殖福利水平，提高了设施养羊场的效益。