基于深度学习模型PANNS-CNN的肉种鸡惊吓声识别研究  

Recognition of Screams of Broiler Breeders Based on Deep Leaing Model PANNS-CNN

作　　者：祝鹏飞 滕光辉 刘健 陈威 吴锦辉 ZHU Pengfei;TENG Guanghui;LIU Jian;CHEN Wei;WU Jinhui(College of Water Resources and Civil Engineering,China Agricultural University,Beijing 100083;Key Laboratory of Agricultural Engineering in Structure and Environment,Ministry of Agriculture and Rural Affairs,Beijing 100083)

机构地区：[1]中国农业大学水利与土木工程学院,北京100083 [2]农业农村部设施农业工程重点实验室,北京100083

出　　处：《中国家禽》2023年第11期105-111,共7页China Poultry

基　　金：山东省重点研发计划(乡村振兴科技创新提振行动计划)(2022TZXD0015);2021年中国农业大学研究生教改研究项目(JG202123)。

摘　　要：为探究鸡舍内行车运动过程中产生的声响对15周龄肉种母鸡惊吓情况,实现惊吓声的快速准确识别与监测,进而在一定程度上提升家禽福利养殖数字化、智能化监测能力,研究以15周龄爱拔益加父母代种母鸡为研究对象,利用在声音识别领域成熟的预训练深度学习模型(PANNS-CNN6、PANNS-CNN10、PANNS-CNN14)对15周龄肉种母鸡转群入产蛋舍情景下因行车运动的惊吓声、行车运动产生的噪声、该品种15周龄肉种母鸡正常鸡鸣声进行识别,并比较三种模型的性能及识别效率。结果显示:在上述3个深度学习模型中PANNS-CNN10的模型综合性能最优,对于验证集的平均准确率达99.81%,对于包含957个样本的验证集识别耗时为3.27s。研究表明,基于PANNS-CNN10构建的肉种鸡,更适合用于实际家禽养殖生产过程中对鸡声音的实时监测,结果在一定程度上弥补了在鸡声音识别领域对于鸡情绪声识别的不足之处,为构建家禽福利养殖体系提供了一定的技术支持。In order to investigate the scaring situation of 15-week-age broiler breeder hens by the sound generated during the movement of traveling vehicles in the chicken coop,to realize the fast and accurate recognition and monitoring of scaring sound,and then improve the digital and intelligent monitoring capability of poultry welfare breeding to a certain extent,15-week-age Arbor Acres Plus S parent breeder hens were used as the study object,and pre-trained deep learning models that were well established in the field of sound recognition(PANNS-CNN6,PANNS-CNN10 and PANNS-CNN14)were used to recognize the scaring sound of 15-week-age broiler breeder hens due to traveling movement and the noise generated by traveling movement when the flock was transfering to laying house,the normal chicken sounds of 15-week-age broiler breeder hens,and the perfor-mance and recognition efficiency of the three models were compared.The results showed that the comprehensive performance of the model of PANNS-CNN10 was the best among the above three deep learning models,with an average accuracy of 99.81%for the verification set and a recognition time of 3.27 s for the verification set containing 957 samples.The study indicated that the meat breeder constructed based on PANNS-CNN10 was more suitable for the sound of chickens in the actual poultry breeding production process real-time monitoring,and the study compensated the deficiency in the field of chicken sound recognition for chicken emotional sound recognition to a certain extent,and provided some technical support for the construction of poultry wel-fare breeding system.

关 键 词：声音识别 PANNS-CNN 肉种鸡 转群 

分 类 号：S831.4[农业科学—畜牧学] TP391.4[农业科学—畜牧兽医]