| Title |
Poultry Vocalization Anomaly Detection via CNN-Autoencoder Hybrid Model |
| Authors |
김나현(Na-Hyeon Kim) ; 김희서(Hee-Seo Kim) ; 이규아(Kyu-A Lee) ; 홍채령(Chae-Ryoung Hong) ; 정경용(Kyoung-Yong Chung) |
| DOI |
https://doi.org/10.5370/KIEE.2026.75.9.2216 |
| Keywords |
Contrastive Autoencoder; CNN; Dual-Path Architecture; Mel-Spectrogram; Poultry Vocalization |
| Abstract |
This paper proposes a hybrid Dual-Path architecture for acoustic anomaly detection in poultry environments, integrating an EfficientNet- B0-based CNN multi-label classifier with an unsupervised contrastive autoencoder (CAE). The CNN classifier performs multi-label classification of three abnormal vocalization types?distress, respiratory disease, and stress?while the CAE suppresses false alarms via reconstruction error, with only partial sensitivity to unseen anomalies. A seven-stage preprocessing pipeline was applied to improve the effective energy ratio from 32.6% to 58.3%. Trained on 70,001 samples, the system achieves a Composite Score of 0.8417, Hamming Accuracy of 0.9459, and Healthy Rejection Rate of 0.8154, with F1-Scores of 0.9969 and 0.9436 for distress and respiratory categories, respectively. |