基于盲源分离与改进BP-DS证据理论的生猪声音多特征融合识别  

Multi-feature fusion recognition of pig sounds based on blind source separation and improved BP-DS evidence theory

作　　者：左义海 张苏楠 贾博 刘乐添 ZUO Yihai;ZHANG Sunan;JIA Bo;LIU Letian(Engineering Training Center,Taiyuan Institute of Technology,Taiyuan 030008,China;Shanxi Animal Husbandry and Veterinary School,Taiyuan 030024,China)

机构地区：[1]太原工业学院工程训练中心,太原030008 [2]山西省畜牧兽医学校,太原030024

出　　处：《黑龙江畜牧兽医》2024年第20期53-62,127,共11页Heilongjiang Animal Science And veterinary Medicine

基　　金：太原工业学院引进人才科研资助项目(2022KJ086)。

摘　　要：为了提高多种声音混合环境下生猪惊骇声、咳嗽声与饥饿声的识别正确率,试验采用基于盲源分离与改进BP-DS证据理论的生猪声音多特征融合识别方法对生猪混合声音识别进行研究。首先,对多个拾音器采集到的混合声音进行能量检测,排除安静环境及微弱环境噪声下声音的干扰,然后利用基于快速独立成分分析(fast independent component analysis,FastICA)的盲源分离方法对采集到的混合声音进行分离;其次,提取分离后声音信号的梅尔倒谱系数及其一阶差分、线性预测倒谱系数及其一阶差分、梅尔倒谱系数与线性预测倒谱系数的组合参数,训练3种特征参数对应的BP神经网络模型;最后,分别利用训练好的BP神经网络模型对生猪声音进行识别,通过改进DS证据理论对识别结果进行融合后输出。结果表明:研究提出的盲源分离方法分离出的生猪声音信号与生猪声音源信号的相似系数均在0.99以上,可以有效地分离生猪混合声音信号;经过多特征融合后正常环境下的识别正确率达到94.22%,较单特征的正确率有一定的提高。说明基于盲源分离与改进BP-DS证据理论的生猪声音多特征融合识别方法可以更高效地识别多种声音混合环境下的生猪声音。In order to improve the recognition accuracy of pigs'fright,cough and hunger sounds in a mixed environment of multiple sounds,a new method of multi-feature fusion recognition of pig's sounds based on blind source separation and improved BP-DS evidence theory was used to study the mixed sound recognition of pigs.Firstly,the energy of mixed sound collected by multiple pickups was detected to eliminate the interference of sound in quiet environment and weak ambient noise.Secondly,blind source separation method based on Fast Independent Component Analysis(FastICA)was used to separate the collected mixed sounds.Thirdly,the mel-frequency cepstral coefficients(MFCC)and their first-order differences(△MFCC),the linear predictive cepstral coefficient(LPCC)and their first-order differences(△LPCC),and the combined parameters of MFCC and LPCC under porcine different states were extracted after blind source separation.The BP neural network models corresponding to three characteristic parameters was trained.Finally,the three trained BP neural network models were applied to recognize pig's sounds,respectively,and the recognition results were fused and output by the improved DS evidence theory.The results showed that the similarity coefficients of the separated porcine sounds and the source sounds by the proposed blind source separation method were all above O.99,which could effectively separate the mixed pig sound signals.After multi-feature fusion,the recognition accuracy in normal environment reached 94.22%,which was higher than that of single feature.The results indicated that the pig sound multi-feature fusion recognition method based on blind source separation and improved BP-DS evidence theory could recognize pig sounds more efficiently in the mixed environment ofmultiple sounds.

关 键 词：声音识别 盲源分离 BP神经网络 改进DS证据理论 多特征融合 

分 类 号：S815[农业科学—畜牧学] TP391.41[农业科学—畜牧兽医]