摘要
以5种猪声为研究对象,首先,用维纳滤波和端点检测对猪声进行预处理,获得有效语料;然后,提取梅尔倒谱系数(Mel Frequency Cepstral Coefficient,MFCC)制作样本集,再构建基于BiLSTM的声学模型学习样本集;最后用训练好的模型对猪声MFCC序列进行分类,实现生猪音频识别。结果表明:(1)通过5折交叉试验验证,5组模型总体识别率均达到90%,最高组为92.52%;(2)用样本集外语料对最优组模型进行算法应用测试,模型对进食、咳嗽、发情、嚎叫和哼叫的样本识别率分别为88.35%、93.65%、90.38%、88.46%、92.63%,总体识别率为90.70%。
Five sounds of Landrace pigs were studied in this paper.First of all,the pigs audio was preprocessed by Wiener filter and endpoint detection to obtain the effective corpus for forming the pig sound sample set.Then,the Mel frequency cepstral coefficients(MFCC)were extracted,which were used to form the pig sound sample set;next,the acoustic model based on BiLSTM was constructed to learn the sample set.Finally,the trained model was used to classify the MFCC sequence of pig sound so that the pig audio was identified.The results showed that(1)Through the 5-fold cross-validation experiment and analysis,the total recognition rates of 5 groups were all over 90%,the highest one is 92.52%.(2)The algorithm application test of the model of the optimal group was carried out with another corpus,the recognition rates of eating samples,cough samples,oestrus samples,scream samples and hum samples were 88.35%,93.65%,90.38%,88.46%and 92.63%,respectively,and the total recognition rate of the model was 90.70%.
作者
邵睿
彭硕
查文文
陈成鹏
辜丽川
焦俊
SHAO Rui;PENG Shuo;ZHA Wen-wen;CHEN Cheng-peng;GU Li-chuan;JIAO Jun(College of Information and Computer,Anhui Agriculture University,Hefei 230036,China)
出处
《合肥学院学报（综合版）》
2022年第2期113-119,共7页
Journal of Hefei University：Comprehensive ED
基金
安徽省科技重大专项“大数据环境下的生猪健康养殖与疫病防控预警关键技术研究”(201903a06020009)
安徽农业大学研究生创新基金项目“基于改进Bi-LSTM的生猪音频分类模型”(2021yjs-52)资助。
