Lev-Ron T, Yitzhaky Y, Halachmi I, Druyan S
School of Electrical and Computer Engineering, Ben-Gurion University of the Negev, 1 Ben Gurion Avenue, P.O.B. 653, Be'er Sheva, 8410501, Israel; Precision Livestock Farming (PLF) Lab, Institute of Agricultural Engineering, Agricultural Research Organization (A.R.O.) - The Volcani Center, 68 Hamaccabim Road, P.O.B 15159 Rishon Lezion, 7505101, Israel; Animal Science Institute, Agricultural Research Organization (A.R.O.) - The Volcani Center, 68 Hamaccabim Road, P.O.B 7505101 Rishon Lezion, 7505101, Israel.
School of Electrical and Computer Engineering, Ben-Gurion University of the Negev, 1 Ben Gurion Avenue, P.O.B. 653, Be'er Sheva, 8410501, Israel.
Animal. 2025 Jan;19(1):101378. doi: 10.1016/j.animal.2024.101378. Epub 2024 Nov 22.
Detecting early-stage stress in broiler farms is crucial for optimising growth rates and animal well-being. This study aims to classify various stress calls in broilers exposed to cold, heat, or wind, using acoustic signal processing and a transformer artificial neural network (ANN). Two consecutive trials were conducted with varying amounts of collected data, and three ANN models with the same architecture but different parameters were examined. The impacts of adding broiler age data as an input attribute and varying input audio waveform lengths on model performance were assessed. Model performance improved with the inclusion of broiler age and longer audio waveforms when trained on smaller datasets. Additionally, the study evaluated the impact of majority vote decision-making across the three ANN model sizes, showing improvement in mean average precision (mAP), particularly for models with shorter audio inputs. Overall, the largest ANN model achieved the highest mAP score of 0.97 for the larger dataset, with small variations among different model sizes. These findings highlight the potential of using a single model to accurately classify multiple types of broiler stress calls. By enhancing the timing of human intervention during critical growth stages, the proposed method may significantly improve broiler welfare and farm management efficiency.
检测肉鸡养殖场的早期应激对优化生长速度和动物健康至关重要。本研究旨在利用声学信号处理和变压器人工神经网络（ANN）对暴露于寒冷、炎热或风中的肉鸡的各种应激叫声进行分类。进行了两项连续试验，收集的数据量不同，并检查了三个架构相同但参数不同的ANN模型。评估了将肉鸡年龄数据作为输入属性以及改变输入音频波形长度对模型性能的影响。在较小数据集上训练时，纳入肉鸡年龄和更长的音频波形可提高模型性能。此外，该研究评估了三种ANN模型规模的多数投票决策的影响，结果显示平均平均精度（mAP）有所提高，尤其是对于音频输入较短的模型。总体而言，最大的ANN模型在较大数据集上实现了最高的mAP分数0.97，不同模型规模之间的差异较小。这些发现凸显了使用单一模型准确分类多种类型肉鸡应激叫声的潜力。通过在关键生长阶段加强人工干预的时机，所提出的方法可能会显著改善肉鸡福利和养殖场管理效率。
