基于深度信念网络的猪咳嗽声识别

摘要

In the early stage,pig cough sound could be detected for early disease warning,and a method based on deep belief nets (DBN) was proposed to construct a pig cough sound recognition model.Pig sounds of Landrace pigs,including cough,sneeze,eating,scream,hum and shaking ears sounds were automatically recorded.The samples were preprocessed by speech enhancement algorithm based on a psychoacoustical model and speech endpoint detection algorithm based on short-time energy to reduce the noise and get useful parts of samples.Based on the dynamic time warping (DTW) algorithm,the shorttime energy characteristics were scaled to a 300-dimensional short-time energy feature vector,while the 24-dimensional MFCC characteristics were scaled to a 720-dimensional MFCC feature vector.And then the 300-dimensional short-time energy feature vector and the 720-dimensional MFCC feature vector were combined to construct a 1020-dimensional vector as the input of the deep belief nets.The number of neuron of the three hidden layers were set to be 42,17 and 7,respectively,so the pig sound recognition model based on DBN was finally designed to be 1020-42-17-7-2.The 5-fold cross validation experiment showed that recognition rate,error recognition rate and total recognition rate of the best experimental group were 94.12％,7.45％ and 93.21％,respectively.Furthermore,the first 479 principal components of 1020 dimcnsion fcature parameters were obtained by PCA dimensionality reduction.The recognition rate,error recognition rate and total recognition rate obtained better performance,and the best experimental group reached 95.80％,6.83％ and 94.29％,respectively.The result demonstrated that the DBN model was effective for the pig cough recognition.%为了在生猪养殖产生呼吸道疾病的初期,通过监测猪咳嗽声进行疾病预警,提出了基于深度信念网络(DBN)对猪咳嗽声进行识别的方法.以长白猪咳嗽、打喷嚏、吃食、尖叫、哼哼、甩耳朵等声音为研究对象,利用基于多窗谱的心理声学语音增强算法和单参数双门限端点检测对猪声音进行预处理,实现猪声音信号的去噪和有效信号检测.基于时间规整算法提取300维短时能量和720维梅尔频率倒谱系数(MFCC)组合成1020维特征参数,将该组合特征参数作为DBN学习和识别数据集,选定3隐层神经元个数分别为42、17和7,构建网络结构为1020-42-17-7-2的5层深度信念网络猪咳嗽声识别模型.通过5折交叉实验验证,基于DBN的猪咳嗽声识别率和总识别率均达到90％以上,误识别率小超过8.07％,最优组猪咳嗽声识别率达到94.12％,误识别率为7.45％,总识别率达到93.21％.进一步基于主成分分析法(PCA)提取1020维特征参数98.01％主成分得到479维特征参数,通过5折交叉实验验证,猪咳嗽声识别率和总识别率相对降维前均有所提高,误识别率有所降低,最优组猪咳嗽声识别率达到95.80％,误识别率为6.83％,总识别率达到94.29％,实验结果表明所建模型是有效可行的.