It is very significant to monitor livestock foraging behavior accurately and in real-time to further improve pasture management and livestock welfare. Although various algorithms have been developed to identify and classify animals' foraging behavior, it still has room to be improved in generality and function. In this study, a representative acoustic dataset generated by typical bodyweight sheep when grazing on various grasses and subsequent ruminating was created. Then an algorithm for the identification and classification of foraging behavior was proposed based on feature extraction technique and deep learning. Specifically, all prominent fragments in the acoustic signal were identified as events by the identification algorithm, and the events were classified as noise, chew, bite, chew-bite, or ruminating behavior through the classification model. The effect of each parameter on the identification algorithm was analyzed, and an optimal set of parameters was derived. As a result, the accuracy of 96.13% was achieved by the identification algorithm with the optimal parameters. Meanwhile, the performances of three common deep network models, including deep neural network (DNN), convolutional neural network (CNN), and recurrent neural network (RNN), were compared. The results showed that the RNN, CNN, and DNN model's accuracy were 93.17%, 92.53%, and 79.43%, respectively. The RNN model had a stronger classification capacity than the CNN model since it involved the inter-dependent information of adjacent events. The CNN model achieved a superior classification performance than the DNN model because the log-scaled Mel-spectrogram representation of the event waveform was more effective than the waveform itself. The algorithm proposed in this study could be well applied to identify and classify all foraging behaviors of typical weight sheep foraging freely on various grasses in the future.

准确实时监测牲畜觅食行为对进一步改善牧场管理和牲畜福利具有重要意义。尽管已经开发了各种算法来识别和分类动物的觅食行为，但在通用性和功能上仍有改进的空间。在这项研究中，创建了典型体重绵羊在吃各种草和随后反刍时生成的代表性声学数据集。然后提出了一种基于特征提取技术和深度学习的觅食行为识别和分类算法。具体而言，通过识别算法将声信号中所有突出片段识别为事件，并将事件分类为噪声、咀嚼、咬合、咀嚼咬合、或通过分类模型反刍行为。分析各参数对识别算法的影响，推导出最优参数集。结果，采用最优参数的识别算法达到了96.13%的准确率。同时，比较了三种常见的深度网络模型，包括深度神经网络（DNN）、卷积神经网络（CNN）和循环神经网络（RNN）的性能。结果表明，RNN、CNN和DNN模型的准确率分别为93.17%、92.53%和79.43%。RNN 模型比 CNN 模型具有更强的分类能力，因为它涉及相邻事件的相互依赖信息。CNN 模型实现了比 DNN 模型优越的分类性能，因为事件波形的对数缩放 Mel 谱表示比波形本身更有效。本研究提出的算法可以很好地应用于未来典型体重绵羊在各种草地上自由觅食的所有觅食行为的识别和分类。