基于声信号特征加权的设施养殖羊行为分类识别

中国西部地区正在发展集约化和规模化的设施养羊业，通过监测羊舍内的声信号可以判别羊只的行为状态，从而为设施养羊的福利化水平评估提取基础依据。梅尔频率倒谱系数（mel frequency cepstrum coefficient，MFCC）模拟了人耳对语音的处理特点且抗噪音性强，被广泛用于畜禽发声信号的特征提取，但其没有考虑各个特征分量表征声信号的能力。该研究构建羊舍无线声音数据采集系统，采集20只羊在设施羊舍内的打斗、饥饿、咳嗽、啃咬和寻伴共5种行为下的声信号，并通过Audacity音频处理软件选出720个清晰且不重叠的声音样本数据。根据MFCC各分量对羊舍声信号表征能力，特征参数提取采用一种熵值加权的MFCC参数，再求其一、二阶差分并进行主成分分析降维，得到优化的19维特征参数。通过对羊舍声信号的声谱图分析，设计了支持向量机二叉树识别模型，并对模型内的4个分类器参数进行网格化寻优测试，该识别模型对羊只5种行为下的声信号进行分类识别，用改进的特征参数与传统MFCC和线性预测倒谱系数（linear predictive cepstrum coefficient，LPCC）进行对比分析。结果表明，该特征参数对5种行为的识别率平均可达83.6%，分别高于MFCC和LPCC参数14.1%和26.8%，羊只打斗和咳嗽行为的声信号属于相似的短时爆发类声音，其识别率分别仅为80.6%和79.5%，啃咬声特征显著不易混淆，其查全率可达到为92.5%，改进特征参数更好的表征了羊舍声信号的特征，提高了羊只不同行为的识别率，为羊只健康和福利状况的监测提供理论依据。

Behavior classification and recognition for facility breeding sheep based on acoustic signal weighted feature

Abstract: Sheep farming husbandry in western region of China has been developing in the manner of intensive and large-scale facility production. Due to the high density of sheep in house, any unusual behavior, such as sheep fighting, will cause a great loss if the sheep farmer is not aware of its happening and takes measures in time. Since the sound from sheep can not only reflect the status of sheep''s health status but also can reflect its response to environment, the behaviors of sheep can be determined by monitoring the sound from the sheep house. This will provide a theoretical basis on evaluation of the welfare level of sheep raising and breeding. In this study, by establishing an audio signal acquisition system for sheep house based on wireless network, the sound signals from 20 sheep under 5 kinds of behaviors (fight, hunger, cough, bite, and search companions) were collected, and then these signals were processed and 720 clear and non-overlapping sound samples were selected in software of Audacity. Although Mel Frequency Cepstrum Coefficients (MFCC) has been widely used for feature extraction of animal sound signals due to its capacities of simulating the processing of speech by human ear and its better noise resistance, it neglects the different contribution of each feature component in characterizing the sound signals from the sheep house. Therefore, in this study, a weighted MFCC method was proposed based on the entropy value method to improve the recognition rate of sheep''s sounds. The weighted MFCC with its first and second order differential was optimized to obtain a 19-dimension feature parameter via principal component analysis. The recognition model of support vector machine binary tree, in which parameters of four classifiers were worked out through grid optimization test, was completed according to the sonograms rendering of five different sheep behaviors. And then, these behaviors were recognized and classified respectively with improved MFCC, traditional MFCC and Linear Prediction Cepstrum Coefficient (LPCC). The results demonstrated that the average recognition rate with the improved MFCC for five different sheep behaviors was up to 83.6%, which was 14.0% and 26.7% higher respectively than MFCC and LPCC. The recognition rate of sheep fight and sheep cough were only 80.6% and 79.5%, respectively, because fight sound and cough sound had similar short outbreak characteristics. The bite sound recall rate was reached to 92.5%, showing that the bite sound was with distinguished feature and uneasy to be confused with other sounds. So, the improved MFCC showed the better performance in characterizing the sounds from sheep house, and in raising the recognition rate of sheep behaviors. Modern techniques of sound analysis have provided tools for analyzing and classifying animal sounds. Taking advantage of this, future bioacoustical research on welfare assessment should focus on comprehensive studies of a broad spectrum of animal specific distress vocalizations. Increasing precise attributions of such utterances to environments, behavioral contexts and relevant physiological parameters will lead to a deeper understanding of their meaning and significance with respect to well-being of farm animals.