Chicken disease detection and localization using multi-noise separation and acoustic recognition

Early detection and prevention of chicken disease are crucial for the sustainability of the poultry industry. However, early identification is often hindered by environmental noise and disease-specific acoustic feature extraction. This study proposes a method for detecting and localization of chicken disease using multi-noise separation and acoustic recognition. A standardized acoustic dataset, including AILT (Avian Infectious Laryngotracheitis), Mycoplasma (Mycoplasma Gallisepticum), Newcastle (Newcastle Disease), and Normal (Health), was established and verified by reverse-transcription Polymerase Chain Reaction (RT-PCR) and PCR test. A fusion model, MSA-BiFPN-EfficientNetV2, was developed, combining bidirectional feature pyramid (BiFPN) for multi-scale acoustic fingerprint semantic collaboration and multi-spectral channel attention (MSA) for enhancing response to pathologically relevant frequency bands. Using log-Mel spectrograms as input, the model performed end-to-end classification of disease-related sounds. Additionally, a localization strategy based on TDOA-GCC-PHAT time delay estimation and spatial grid mapping accurately localized diseased chickens within the cages. Experimental results showed an accuracy of 96.17% and F1-score of 96.28% on the test set, with Mycoplasma and AILT achieving F1-scores of 97.34% and 97.09%, respectively. The model also achieved an average localization error at 0.085 meters. This study presented a novel approach for implementing an acoustic-based early warning system for chicken diseases.