基于轻量化卷积神经网络的笼养鸡异常发声识别方法

Lightweight Convolutional Neural Network-based Method for Abnormal Vocalization Recognition in Cage-housed Chickens

摘要:

针对笼养鸡舍复杂声环境下鸡只异常发声识别易受背景噪声干扰、异常样本稀缺及类别不平衡影响而导致识别稳定性不足的问题,提出一种基于轻量化卷积神经网络(CNN)的鸡只发声三分类识别方法。 以鸡舍采集的 “ 叫声” “ 喷嚏类” 和“ 其他”3 类声音为研究对象,构建包含 691 段音频样本的数据集,按训练集、验证集和测试集进行划分,并将音频转换为 128 × 130 的对数梅尔频谱图作为模型输入。 在此基础上设计轻量级 SimpleCNN 模型,并与Inception V3 和 ResNet-50 进行对比试验。 结果表明:SimpleCNN 在测试集上的总体准确率达到0. 95,“ 叫声” “ 喷嚏类” 和“ 其他”3 类的F1分数分别为0. 95、0. 91 和 0. 98,对少数类“ 喷嚏类” 和噪声类“ 其他” 均表现出较好的识别能力;相比之下,Inception V3 和ResNet 50的总体准确率分别为0. 56和0. 32,且存在明显类别偏置,其中ResNet-50出现类别塌缩现象。混淆矩阵与可解释性分析表明,SimpleCNN 能够更有效地聚焦目标发声事件的关键时频区域,有效降低“ 喷嚏类” 与“ 其他”之间的混淆。研究结果表明,轻量化CNN结构在小样本、强噪声条件下具有较好的判别稳定性和部署潜力,可为鸡舍异常声音监测与非接触式健康预警提供技术支撑。

Abstract:

Aiming to address the limited stability of abnormal vocalization recognition in caged layer houses under complex acoustic conditions, where strong background noise, scarce abnormal samples, and class imbalance pose major challenges, a lightweight convolutional neural network (CNN)for three-class chicken sound classification was proposed. A dataset of 691 audio clips, including calls, sneeze-like sounds, and other sounds, was constructed from recordings collected in commercial hen houses and divided into training, validation, and test sets. Each clip was converted into a 128 × 130 log Mel spectrogram as model input. A lightweight SimpleCNN was developed and compared with Inception V3 and ResNet-50. Results showed that SimpleCNN achieved an overall test accuracy of 0. 95, with F1- scores of 0. 95, 0. 91, and 0. 98 for calls, sneeze-like sounds, and other sounds, respectively, demonstrating strong recognition performance for both the minority abnormal class and the noise class. In contrast, Inception V3 and ResNet 50 achieved overall accuracies of 0. 56 and 0. 32, respectively, and showed clear class bias, with ResNet-50 exhibiting class collapse. Confusion matrix and interpretability analyses further indicated that SimpleCNN more effectively focused on key time-frequency regions of target vocal events and reduced confusion between sneeze-like sounds and other sounds. These findings suggested that lightweight CNNs offered robust discrimination and practical deployment potential under small-sample, high-noise conditions, providing technical support for abnormal sound monitoring and non- contact health warning in layer houses.