基于偏度聚类的哺乳期母猪声音特征提取与分类识别  

Feature Extraction and Classification Based on Skewness Clustering Algorithm for Lactating Sow

作　　者：闫丽[1,2] 邵庆[1] 吴晓梅 谢秋菊[1] 孙昕[4] 韦春波[5] 

机构地区：[1]黑龙江八一农垦大学信息技术学院,大庆163319 [2]南京农业大学工学院,南京210031 [3]国网黑龙江省电力公司黑河供电公司,黑河164300 [4]哈尔滨医科大学基础医学院,哈尔滨150081 [5]黑龙江八一农垦大学动物科技学院,大庆163319

出　　处：《农业机械学报》2016年第5期300-306,共7页Transactions of the Chinese Society for Agricultural Machinery

基　　金：黑龙江省青年科学基金项目(QC2014C078;QC2013C031);黑龙江省教育厅科研项目(12541493);大庆市指导性科技计划项目(szdfy-2015-23)

摘　　要：哺乳期是母猪繁育仔猪的关键时期,哺乳母猪特有的发声是其生理、情绪健康及其对仔猪看护的母性能力的最直接表达。哺乳期间母猪所发声音种类众多,增加了快速定位及准确识别特定声音类型的复杂度,以小梅山母猪的哺乳声、饮水声、采食声及无食咀嚼声等常见声音为研究对象,以功率比作为特征向量,对频域进行更精细的能量计算,提出基于偏度的子带聚类法合并特征不显著的子带,减少特征向量数量,构建支持向量机(SVM)的声音分类识别器,统计各类声音的发声时长;进一步以单个哺乳周期为对象,建立成功哺乳的声音模式。试验结果表明,哺乳声、无食咀嚼声、采食声和饮水声的最大功率比分别位于[0 Hz,1 000 Hz]、[1 000 Hz,1 500 Hz]、[1 500 Hz,2 500 Hz]和[2 500 Hz,8 000 Hz]子带内,以4个子带的功率比为特征的声音判别模型的识别率分别为100%、100%、95.17%、96.61%,与等间隔子带划分及主成分分析法比较,减少了特征向量的数量,且显著提高了识别算法的精度,进一步应用在母猪分娩舍内,实现了对哺乳母猪的母性能力及其健康状况的无应激、实时监测。The lactation period is a critical period for sows to breed their piglets,and the specific voice of lactating sows in this period is the most direct expression of their physiology,emotional health,and maternal ability to care for piglets. The rapid location and accurate identification will be more complex due to a variety of vocalizations during this period. Therefore,the vocalizations of nursing grunt,drinking,feeding and sham chewing were observed,and a fine energy calculation for frequency domain with a power ratio as a vector was carried out. Then,the sub-band clustering method based on skewness was presented to merge the sub bands without significant characteristics to reduce the number of parameters. Thirdly,the recognizer for sow ＇s vocalizations was built based on support vector machine（ SVM） to calculate the duration of the different types of vocalization. A sound mode of successful nursing was established further within single lactation circle. It is shown that the max power ratio frequencydomain of the nursing grunt,the sham chawing,the feeding and the drinking are ranged from 0 Hz to1 000 Hz,1 000 Hz to 1 500 Hz,1 500 Hz to 2 500 Hz,and 2 500 Hz to 8 000 Hz,respectively. The accuracy of the vocalization recognition mode with these four sub bands power ratio frequency as parameters were 100%,100%,95. 17% and 96. 61%,respectively. Compared with the uniformlyspaced sub-band division and principal component analysis（ PCA）,the number of features was reduced,and the recognition accuracy was significantly improved in the clustering algorithm based on skewness.Thus,the proposed method could be further applied in the health and maternal ability of sows monitoring real-timely and nonstressly.

关 键 词：哺乳母猪 声音识别 偏度 聚类 降维 哺乳模式 

分 类 号：S126[农业科学—农业基础科学]