【作者】 唐杰；

【导师】 陆明洲；

【作者基本信息】 南京农业大学 ， 电子信息（专业学位）， 2022， 硕士

【摘要】 近年来,国民经济持续保持长足发展,农业经济方面的效益是其中的重要一环,而畜牧业作为中国农业增效、增收的重要产业也随之快速发展,所以畜牧养殖规模化成为必要的养殖模式。妊娠期内母羊的健康状况和分娩工作的好坏,直接关系到养殖户的收入。母羊分娩属于一个正常的生理过程,一般不需人工助产,然而,小羊羔出生后,养殖人员应立即清除羔羊口、鼻、耳中的粘液,以防止异物引起的肺炎或误吞羊水而窒息。也应当时刻注意观察,如果发现胎衣排出,马上将胎衣拿走,防止被母羊吞食后养成母羊咬羔、吃羔等恶癖。因此,母羊临产预警在养殖生产实践中具有较好的应用价值,对羊羔的成活率及健康水平有重要意义。目前对于孕羊临产的监测主要依赖于具有丰富养殖经验的技术员人工巡查判断,人工巡查方式难以实现对每一头孕羊都进行实时连续观测及其临产状态的准确判断,因夜间人工巡查不连续而导致的深夜产羔的孕羊及其羔羊缺乏助产、照料的情况时有发生。虽然也有一些利用接触式传感器(如三轴加速度传感器)获取母羊产前行为数据并做临产预测的研究,但这些研究提出的方法一般只是利用产前母羊姿态特征,临产预测性能仍有较大的提升空间。录音笔因其采音准确、便于携带等特点,越来越多地被用作监测农场动物生理和行为,相应的基于音频处理技术的动物行为自动分析也成为智慧畜牧技术领域的热点研究方向之一。本文针对母羊临近分娩时会发出努责声音的特点,采用录音笔采集母羊产生的声音信号,训练自动识别母羊产前努责声的深度学习模型。模型测试结果表明,音节分割后的声学分类模型能以97.32%的平均准确率检测母羊努责声音。本文主要研究内容如下:(1)音频预处理与特征提取。音频预处理过程主要包括去噪、分帧加窗以及音节分割等。课题中设计提取噪声中出现频率最高的风扇声作为加性噪声,使用谱减法将原始带噪音频减去加性噪声,进行降噪处理。在对语音特征中常用的特征参数LPCC与MFCC对比后,确定MFCC作为特征输入被送入到识别模型中进行学习、训练。本文对努责声、咀嚼声、其它声音(如撞击栏杆声)三类非静默声音信号进行时域分析后,发现仅应用时域信息难以将其区分。为了得到连续的努责音频,采用音节分割算法,通过能零比(EZR)、峰谷点等特征,将连续音频以静默端自动分割成多段非静默音频,并将非静默段人工标注为努责声“1”、咀嚼声“2”、其它声音“3”。(2)母羊努责音频分类识别算法研究。将母羊产羔前一周内努责、咀嚼、其它声音数据集划分为训练集和测试集,分别提取努责、咀嚼、其它声音的MFCC特征系数,再分别将音节分割前与音节分割后的音频数据中所提取的MFCC特征参数送入LSTM-CTC和BRNN-CTC完成分类模型训练。使用测试集数据验证了两个识别模型的性能,并将准确率和召回率作为模型的评价值。结果表明音节分割后的BRNN-CTC对母羊努责检测识别准确率表现最高,其平均检测准确率为97.32%。(3)母羊产前努责声自动识别模型测试。通过BRNN-CTC模型识别母羊连续努责音频,实现努责频次的自动统计。绘制母羊产前努责频次折线图并分析母羊生产过程中努责频次特点,实验结果表明顺产羊在临近生产前努责频次会有上升的趋势,难产母羊在生产时努责频次没有呈现一个逐渐上升的趋势,且难产母羊努责声音持续时间明显长于顺产母羊。本文提出了一种以音节分割后母羊产前声学信号为输入的母羊努责音频识别BRNN-CTC模型,实现努责频次的自动统计。本文构建的母羊努责声识别模型可作为母羊临近分娩预警系统的基础,也可用于难产母羊的自动识别。为降低人工巡查母羊是否临产的人力、时间开销,提高羊羔存活率,提供了技术与方法基础。更多还原

【Abstract】 In recent years,the national economy has maintained rapid development,and the benefit of agricultural economy is an important part of it.As an important industry to increase the efficiency and income of China’s agriculture,animal husbandry has also developed rapidly,so the large-scale animal husbandry has become a necessary breeding mode.The health status of ewe during gestation and the quality of delivery work are directly related to the income of farmers.Ewe childbirth is a normal physiological process and generally does not need artificial assistance.However,after the birth of lambs,breeders should immediately remove the mucus in lambs’ mouths,noses and ears to prevent pneumonia caused by foreign bodies or asphyxia caused by swallowing amniotic fluid.Also should always pay attention to observation,if it is found that the afterbirth is discharged,immediately take away the afterbirth,to prevent being swallowed by ewe to develop ewe bite the lamb,eating the lamb and other evil habits.Therefore,the ewe birth warning has a good application value in breeding practice,and is of great significance to the survival rate and health level of lambs.For pregnant sheep and labor monitoring rely mainly on judging has rich breeding experience technicians artificial inspections,artificial search way is difficult to implement for every pregnant sheep,accurate judgment the real-time continuous observations and its status in labor,caused by artificial patrol at night discontinuous lambing late pregnant sheep and lambs lack of midwifery,take care.Although there are some studies that use contact sensors(such as triaxial acceleration sensors)to obtain prenatal behavior data of ewes and make labor prediction,the methods proposed in these studies generally only use prenatal posture characteristics of ewes,and there is still a large space for improvement in labor prediction performance.The recording pen is more and more used to monitor the physiology and behavior of farm animals due to its characteristics of accurate sound collection and portability.The corresponding automatic analysis of animal behavior based on audio processing technology has become one of the hot research directions in the field of intelligent animal husbandry technology.In this paper,we used a recording pen to collect the sound signal generated by ewe sheep,and trained a deep learning model to automatically recognize the sound of ewe sheep before delivery.The model test results show that the acoustic classification model can detect the ewe nuances with an average accuracy of 97.32%.The main research contents of this paper are as follows:(1)Audio preprocessing and feature extraction.The process of audio preprocessing mainly includes denoising,frame segmentation and syllable segmentation.In the subject,the fan sound with the highest frequency is extracted as additive noise,and the original band noise frequency is subtracted from the additive noise by spectral subtraction.After comparing LPCC and MFCC,which are commonly used in speech features,MFCC is determined to be sent into the recognition model as feature input for learning and training.In this paper,the time domain analysis of three kinds of non-silent sound signals,such as squealing sound,chewing sound and other sounds(such as the sound of hitting the railings)is carried out,and it is found that it is difficult to distinguish them only by using time domain information.In order to obtain continuous nuze audio,a syllable segmentation algorithm was used to automatically divide the continuous audio into several non-silent audio segments with the silent end by the characteristics of zero energy ratio(EZR)and peak and valley points,and the non-silent segments were manually marked as nuze "1",chewing "2" and other sounds "3".(2)Study on ewe nuze audio classification and recognition algorithm.The data set of nuzing,chewing and other sounds in the week before lambing were divided into training set and test set,and the MFCC characteristic coefficients of nuzing,chewing and other sounds were extracted respectively.The MFCC feature parameters extracted from the audio data before and after syllable segmentation were sent to LSTM-CTC and BRNN-CTC respectively to complete the classification model training.Test data were used to verify the performance of the two recognition models,and the accuracy and recall rates were taken as the evaluation values of the models.The results showed that BRNN-CTC had the highest recognition accuracy of ewe nuze detection after syllable segmentation,with an average detection accuracy of 97.32%.(3)Test of automatic recognition model of ewe nuze noise before delivery.Using BRNN-CTC model to recognize ewe’s continuous nuzee frequency,the automatic statistics of nuze frequency was realized.Sheep ewes prenatal nuze frequency line chart and analysis in the process of production nuze frequency characteristics,the experimental results show that natural birth sheep near nuze frequency before production there will be a rising trend,dystocia ewes during production nuze frequency did not present a rising trend,and dystocia ewes nuze sound lasts longer than natural labor ewes.In this paper,a BRNN-CTC model for ewe nuze audio recognition based on prenatal acoustic signals of ewe after syllable segmentation was proposed to realize automatic statistics of nuze frequency.The model can be used as the basis of the early warning system of ewe near delivery and also can be used for the automatic recognition of ewe with difficult delivery.It provides a technical and methodological basis for reducing the labor and time cost of manual inspection of ewe and improving the survival rate of lambs.更多还原