深层神经网络在家畜反刍行为识别上的应用

Application of Deep Neural Network on the Livestock Ruminating Behaviour Recognition

针对反刍家畜的反刍行为结合前馈型神经网络提出了一个简单的方法,自动分析和识别反刍家畜的反刍事件。用声音信号表示反刍事件的候选项,然后用前馈神经网络模型对这些候选项的特征进行学习,通过训练好的模型将这些侯选项识别为反刍事件和非反刍事件。测试用的声音信号总共24 min,分别来自三只湖羊,喂给这些湖羊的饲料是典型的混合饲料。为了充分测试提出的方法,测试数据分别使用了包含较少噪声和包含较多噪声的数据,并且更加注重后者,因为噪音在所有的声学模型中都是一个很大的挑战。实验表明,该方法获得的湖羊反刍事件的次数和人工统计的反刍次数较为接近,统计次数的正确率超过了90%,并且所有反刍事件的正确匹配率约为87%。

Combined forward neural network and the ruminating behaviour of ruminants,a simple method was presented to analysis and recognize ruminant events automatically. Sound signals were used to indicate the ruminant event candidate,and then the deep neural network model learn the characteristics of these candidates,through the trained model these candidates are identified as ruminant events and non-ruminant events. A total of 24 minutes sound data were used to test the method,these sound signals respectively from three lake sheep. The food feed to these lake sheep are typical of the mixed feed. In order to fully test our method,the test data contain less noise and some contain more noise are used,what＇s more,more attention were paid to the latter,because the noise is a big challenge for all acoustic models. Experiment shows that the number of ruminant events given by the model is closer to the labor statistics,correct recognition more than 90% and the matching rate over 87% of ruminant events.