Visual monitoring of behaviour on-farm is mostly challenging due to the number of animals to be observed and the time required. However, behavioural problems such as cannibalism in turkeys may be preceded by subtle changes in behaviour. Machine learning techniques allow automatic behavioural monitoring of livestock to be carried out under different farming conditions. The aim of this study was to develop and test a novel pecking activity detection tool for potential use on turkey farms by means of acoustic data and a convolutional neural networks (CNN) model. Under near to commercial conditions, two metallic balls were used as pecking objects and suspended from the ceiling. Each pecking object was equipped with a microphone connected via a cable to a top view camera positioned on the ceiling. The recorded sound data were sampled in slots of 1 s and high pass filtering was performed to eliminate background noises. A total of 9200 filtered sound files were used for training and validation, and 3900 for testing set. They were labelled manually as peck or non-peck, using 7360 (80%) for training and 1840 (20%) for validation files, and fed into the CNN model. An additional 3900 new filtered sound clips were used to test the detection phase of the trained model. The experimental results illustrate that the deep learning-based detection method achieved high overall accuracy, precision, recall and F1-score of 96.8, 89.6, 92.0 and 90.8% in the detection phase. This indicates that the proposed technique could be used as a precise method for the detection of pecking activity levels in turkeys.

中文翻译：

---

使用声学数据和深度学习技术检测群养火鸡的啄食活动

由于要观察的动物数量和所需的时间，对农场行为的视觉监测主要具有挑战性。然而，行为问题（如火鸡自相残杀）之前可能会出现行为的细微变化。机器学习技术允许在不同的耕作条件下对牲畜进行自动行为监测。本研究的目的是通过声学数据和卷积神经网络 (CNN) 模型开发和测试一种新型啄食活动检测工具，可用于火鸡农场。在接近商业条件下，两个金属球被用作啄食物体并悬挂在天花板上。每个啄食对象都配备了一个麦克风，麦克风通过电缆连接到位于天花板上的顶视摄像头。记录的声音数据以 1 s 的时隙进行采样，并执行高通滤波以消除背景噪声。总共 9200 个过滤后的声音文件用于训练和验证，3900 个用于测试集。它们被手动标记为啄食或非啄食，使用 7360 (80%) 用于训练和 1840 (20%) 用于验证文件，并输入 CNN 模型。另外 3900 个新的过滤声音片段用于测试训练模型的检测阶段。实验结果表明，基于深度学习的检测方法在检测阶段实现了96.8%、89.6%、92.0%和90.8%的高整体准确率、准确率、召回率和F1-score。这表明所提出的技术可用作检测火鸡啄食活动水平的精确方法。