Abstract As the poultry farming industry attaches great importance to the welfare of poultry, an understanding of the daily behaviours of poultry is becoming more important. Eating behaviour is an important daily behaviour in poultry. Based on the difference between the eating vocalization and the normal vocalization of poultry, this study proposed an automatic detection method for the eating behaviour of poultry based on an audio analysis and a time sequence model. First, the short time energy (STE) and short time zero crossing rates (STZ) of poultry vocalizations and environmental sounds were analysed, and a four-threshold poultry vocalization automatic selection method based on the STE and STZ was proposed. Then, these poultry vocalizations were characterized according to variations in time. Three types of poultry vocalization network (PV-net) were proposed for classify the eating vocalization and the normal vocalization of poultry: PV-net1, PV-net2 and PV-net3. In the experiments, vocalization of 18 chickens was collected. Results showed that the recognition rates of PV-net1, PV-net2 and PV-net3 were 93.5%, 94.5% and 96%, respectively; their sensitivities were 96%, 97% and 96%, respectively; and their specificities were 91%, 92% and 96%, respectively. This method is beneficial to the protection of poultry welfare and will be significant for use in monitoring and controlling the eating behaviour of poultry in the poultry farming industry.

中文翻译：

---

基于发声信号的家禽进食行为智能检测方法

摘要 随着家禽养殖业对家禽福利的高度重视，了解家禽的日常行为变得越来越重要。采食行为是家禽重要的日常行为。本研究基于家禽进食发声与正常发声的差异，提出了一种基于音频分析和时序模型的家禽进食行为自动检测方法。首先，分析了家禽发声和环境声的短时能量（STE）和短时过零率（STZ），提出了一种基于STE和STZ的四阈值家禽发声自动选择方法。然后，根据时间的变化对这些家禽的发声进行了表征。提出了三种类型的家禽发声网络（PV-net）来对家禽的进食发声和正常发声进行分类：PV-net1、PV-net2和PV-net3。在实验中，收集了 18 只鸡的发声。结果表明，PV-net1、PV-net2和PV-net3的识别率分别为93.5%、94.5%和96%；它们的灵敏度分别为 96%、97% 和 96%；它们的特异性分别为 91%、92% 和 96%。该方法有利于保护家禽福利，对家禽养殖业中家禽采食行为的监测和控制具有重要意义。结果表明，PV-net1、PV-net2和PV-net3的识别率分别为93.5%、94.5%和96%；它们的灵敏度分别为 96%、97% 和 96%；它们的特异性分别为 91%、92% 和 96%。该方法有利于保护家禽福利，对家禽养殖业中家禽采食行为的监测和控制具有重要意义。结果表明，PV-net1、PV-net2和PV-net3的识别率分别为93.5%、94.5%和96%；它们的灵敏度分别为 96%、97% 和 96%；它们的特异性分别为 91%、92% 和 96%。该方法有利于保护家禽福利，对家禽养殖业中家禽采食行为的监测和控制具有重要意义。